0% found this document useful (0 votes)
193 views

Chatterjee P. Machine Learning Algorithms and Applications in Engineering 2023

Uploaded by

bryam vargas
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
193 views

Chatterjee P. Machine Learning Algorithms and Applications in Engineering 2023

Uploaded by

bryam vargas
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 339

Machine Learning Algorithms

and Applications in
Engineering

Machine Learning (ML) is a subfield of artificial intelligence that uses soft


computing and algorithms to enable computers to learn on their own and
identify patterns in observed data, build models that explain the world, and
predict things without having explicit pre-​programmed rules and models.
This book discusses various applications of ML in engineering fields and the
use of ML algorithms in solving challenging engineering problems ranging
from biomedical, transport, supply chain and logistics, to manufacturing
and industrial. Through numerous case studies, it will assist researchers and
practitioners in selecting the correct options and strategies for managing
organizational tasks.

Features:

• Includes the latest research contributions on ML perspectives and


applications.
• Follows an algorithmic approach for data analysis in ML and includes
real-​time engineering case studies.
• Addresses the emerging issues in computing such as deep learning,
Internet of Things, and data analytics.
• Focuses on ML techniques that are unsupervised and semi-​supervised
for unseen and seen data sets on commercial value-​added research
applications.
• Discusses how the advancements on data science and computing
domain open possibilities for cross-​disciplinary connections.

This book will benefit professionals, academic researchers, and graduate


students in computer science and engineering, biomedical, mechanical, elec-
trical, transport, manufacturing, and industrial engineering.
Machine Learning
Algorithms and
Applications in
Engineering

Edited by
Prasenjit Chatterjee, Morteza Yazdani,
Francisco Fernández-​Navarro, and
Javier Pérez-​Rodríguez
First edition published 2023
by CRC Press
6000 Broken Sound Parkway NW, Suite 300, Boca Raton, FL 33487-​2742
and by CRC Press
4 Park Square, Milton Park, Abingdon, Oxon, OX14 4RN
CRC Press is an imprint of Taylor & Francis Group, LLC
© 2023 Taylor & Francis Group, LLC
Reasonable efforts have been made to publish reliable data and information, but the author
and publisher cannot assume responsibility for the validity of all materials or the consequences
of their use. The authors and publishers have attempted to trace the copyright holders of all
material reproduced in this publication and apologize to copyright holders if permission to
publish in this form has not been obtained. If any copyright material has not been acknowledged
please write and let us know so we may rectify in any future reprint.
Except as permitted under U.S. Copyright Law, no part of this book may be reprinted,
reproduced, transmitted, or utilized in any form by any electronic, mechanical, or other means,
now known or hereafter invented, including photocopying, microfilming, and recording, or in
any information storage or retrieval system, without written permission from the publishers.
For permission to photocopy or use material electronically from this work, access www.
copyri​ght.com or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive,
Danvers, MA 01923, 978-​750-​8400. For works that are not available on CCC please contact
[email protected]
Trademark notice: Product or corporate names may be trademarks or registered trademarks and
are used only for identification and explanation without intent to infringe.
ISBN: 9780367569129 (hbk)
ISBN: 9780367612559 (pbk)
ISBN: 9781003104858 (ebk)
DOI: 10.1201/​9781003104858
Typeset in Palatino
by Newgen Publishing UK
Contents

Preface......................................................................................................................vii
Organization of the Book....................................................................................... ix
The Editors.............................................................................................................xiii

1 Machine Learning for Smart Health Care................................................... 1


Rehab A. Rayan

2 Predictive Analysis for Flood Risk Mapping Utilizing


Machine Learning Approach....................................................................... 17
Aditya Singh, Sunil Khatri, Sandhya Save, and Hemant Kasturiwale

3 Machine Learning for Risk Analysis......................................................... 35


Parita Jain, Puneet Kumar Aggarwal, Kshirja Makar, Riya Garg,
Jaya Mehta, and Poorvi Chaudhary

4 Machine Learning Techniques Enabled Electric Vehicle....................... 55


Shyamalagowri Murugesan and Revathy Jayabaskar

5 A Comparative Analysis of Established Techniques and


Their Applications in the Field of Gesture Detection............................ 73
Muskan Jindal, Eshan Bajal, and Shilpi Sharma

6 Brain–Computer Interface for Dream Visualization using


Deep Learning................................................................................................ 93
Brijesh K. Soni and Akhilesh A. Waoo

7 Machine Learning and Data Analysis Based Breast Cancer


Classification..................................................................................................117
Souvik Das, Rama Chaitanya Karanam, Obilisetty Bala Krishna, and
Jhareswar Maiti

8 Accurate Automatic Functional Recognition of Proteins:


Overview and Current Computational Challenges.............................. 131
Javier Pérez-Rodríguez, Morteza Yazdani, and Prasenjit Chatterjee

9 Taxonomy of Shilling Attack Detection Techniques in


Recommender System................................................................................. 141
Abhishek Majumder, Keya Chowdhury, and Joy Lal Sarkar

v
vi Contents

10 Machine Learning Applications in Real-World Time


Series Problems............................................................................................ 161
Antonio Manuel Durán-Rosal and David Guijo-Rubio

11 Prediction of Selective Laser Sintering Part Quality Using


Deep Learning.............................................................................................. 177
Lokesh Kumar Saxena and Pramod Kumar Jain

12 CBPP: An Efficient Algorithm for Privacy-Preserving


Data Publishing of 1:M Micro Data with Multiple Sensitive
Attributes....................................................................................................... 195
Jayapradha Jayaram, Prakash Manickam, Apoorva Gupta, and
Madhuri Rudrabhatla

13 Classification of Network Traffic on ISP Link and Analysis of


Network Bandwidth during COVID-19.................................................. 213
V. Ajantha Devi, Yogendra Malgundkar, and Bandana Mahapatra

14 Integration of AI/Ml in 5G Technology toward Intelligent


Connectivity, Security, and Challenges................................................... 239
Devasis Pradhan, Prasanna Kumar Sahu, Rajeswari, Hla Myo Tun,
and Naw Khu Say Wah

15 Electrical Price Prediction using Machine Learning Algorithms....... 255


Swastik Mishra, Kanika Prasad, and Anand Mukut Tigga

16 Machine Learning Application to Predict the Degradation


Rate of Biomedical Implants..................................................................... 271
Pradeep Bedi, Shyam Bihari Goyal, Prasenjit Chatterjee,
and Jugnesh Kumar

17 Predicting the Outcomes of Myocardial Infarction Using


Neural Decision Forest............................................................................... 283
Akashdeep Singh Chaudhary and Ravinder Saini

18 Image Classification Using Contrastive Learning................................. 297


Abhyuday Trivedi, Anjali Hembrom, Arkajit Saha, Tahreem Fatima,
Shreya Dey, Monideepa Roy, and Sujoy Datta

Index...................................................................................................................... 315
Preface

Some of the decisions taken in engineering, from bioinformatics to risk ana-


lysis, are supported by information obtained from data. In recent decades,
both academic and private institutions have sought to agglomerate that
information in data centers, aiming to extract representative knowledge out
of it. In particular, the improvement in information-​storage resources has
reached a point where almost every institution may store and preserve a
considerable amount of data related to its daily services. As expected, raw
data stored in those data centers is useless and requires analysis through
advanced algorithms. During the twenty-​first century, these algorithms have
had a remarkable development, outperforming classical statistical methods.
Machine learning (ML) and big data technologies have become the solution
to massive information processing needs that exceed the capabilities of con-
ventional computing. They are characterized by three features: volume (a
large amount of data is produced), velocity (the speed at which these data
are obtained and accessed), and variety (data diversity as they come from
different sources; e.g., structured, unstructured, multimedia, textual). In this
book, we will present several novel applications of ML approaches to engin-
eering problems extremely relevant in practice, ranging from health care to
risk analysis.

vii
Organization of the Book

This book is organized into 18 chapters. A brief description of each chapter is


presented below.
Chapter 1 reports successful ML models in smart health care through clin-
ical applications such as medical diagnostics and precision or monitoring
health, their potential, and limitations. With the increasing power of
supercomputers, ML developments would have the proper environment for
managing health such as detecting diseases precisely and early, improving
diagnosis, and better therapies.
Chapter 2 proposes predictive ML models for assisting exact foreseeing
models in recognizing and mapping flood hazard territories. The calculations
obtained through ML techniques lead to hazard exhaustion, arrangement
suggestion, alleviating loss of helpful lives, and decreasing flood-​related
property harm.
The main purpose of Chapter 3 is to identify the flaws in current air quality
and to recommend measures or policies to improve it so that it is safe to con-
sume. Because each country employs a different scale for evaluation, the data
provided by the Central Pollution Control Board of India are utilised to ana-
lyse the level of air in the zone under investigation.
Chapter 4 implements an expert system model for solving the problem
associated with the precise prediction of the dynamic trajectory of an autono-
mous vehicle. This was accomplished by deriving a new equation for deter-
mining the lateral tire forces and adjusting some of the vehicle parameters
under road test conductions. A universal approach to performing the reverse
engineering of electric power steering (EPS) for external control is also
presented in this chapter.
Chapter 5 proposes a speech recognition model that compiles contemporary
expertise on the detection of gestures and their associated mannerisms while
proposing modifications for inclusivity of a wider audience, with a prom-
inent emphasis on people with communication disabilities. Furthermore, the
work done in this chapter may culminate in the development of real-​world
applications such as determining the attentivity of students in real time.
Chapter 6 covers the fundamental phenomenon of brain–​computer inter-
face as an emerging technology that may help to become possible to com-
municate dreamers into their real dream through modern technologies. This
chapter contains only the essential concepts and theories in integration for
developing a deep learning-​based model for mapping dream contents. The
first part of the chapter narrates dream theories and related cortical regions,
whereas the second one concentrates briefly on pillar technology brain–​
computer interface.

ix
x Organization of the Book

Chapter 7 proposes the use of ML algorithms to classify breast cancer more


efficiently through the exploitation of existing data on cancer. This chapter
examines the cancer patient data obtained from images, how the data is
preprocessed, how different ML models are trained, and chooses the best-​
trained model for the classification of breast cancer. The current study helps
not only in identifying the best ML model to classify breast cancer but also to
formulate a process of diagnosing breast cancer based on the preprocessed
data obtained from the raw images of breast cancer.
Chapter 8 focuses on the prediction of protein functions, addressing the
problem from a computer science (machine learning-​based) perspective.
Indeed, it can be seen as an overview on the topic of the prediction of
protein function aimed especially at computer scientists and researchers
who work on machine learning. It will attempt to formally define the
multilabel problem of protein function prediction, to explain its magni-
tude and importance, and why it remains as an open and challenging
problem nowadays.
In Chapter 9, the profile injection attack methods, collaborative filtering in
recommender systems, and shilling attack detection schemes are discussed
from an ML perspective. The shilling attack detection technique works on
two parameters, namely, rating parameter and rating and time interval
parameter. Based on the working parameters, the shilling attack detection
techniques have been classified. They have also been classified based on their
output. A brief discussion on every scheme has also been presented.
Chapter 10 highlights some of the primary time series data mining (TSDM)
tasks, among which time series preprocessing, segmentation, or prediction
are widespread in the literature. ML techniques along with other time series
approaches are detailed throughout a set of real-​world applications, such
as the wave height time series reconstruction, the detection, and prediction
of tipping points in paleoclimatology time series, the forecasting of low-​
visibility events produced by the existence of fog or convective situations,
among others.
Chapter 11 uses ML algorithm for predicting quality of a product in terms
of minimum shrinkage ratio produced by selective laser sintering (SLS) using
some key parameters including surrounding working temperature, laser
scanning speed, layer thickness, scanning mode, hatch distance, laser power,
and interval time.
Chapter 12 proposes class batch pid partitioning (CBPP) algorithm to pro-
vide an optimal balance between utility and privacy for data publishing of
1:M micro data with multiple sensitive attributes.
Chapter 13 focuses on assessing the performance data obtained from
real networks during the COVID-​19 pandemic using scripts and existing
applications for monitoring apps and networks in real time.
Chapter 14 highlights the integration of AI/​ML for 5G network considering
different constraints.
Organization of the Book xi

In Chapter 15, ML methods are used to forecast power prices using an


open-​source dataset. For better prediction and visualization of electricity
prices than the already anticipated values in the dataset, a web tool called
Jupyter notebook is employed.
In Chapter 16, ML is used to forecast the qualities of various materials as
well as their deterioration rates. This model also generates the composition of
materials, as well as their mechanical and biological properties, saving time
and reducing failures during experimental examination.
The objective of Chapter 17 is to predict the various outcomes of Myocardial
Infarction (commonly known as a heart attack) using Neural Decision Forest.
The goal of Chapter 18 is to show a current implementation of SimCLR, a
basic framework for contrastive learning of visual representations, which is a
contrastive self-​supervised learning method. This research looks into how to
learn these visual representations by examining the framework's most critical
sections on a regular basis.
The Editors

Prasenjit Chatterjee is currently the Dean (Research and Consultancy) at


MCKV Institute of Engineering, West Bengal, India. He has published over
120 research papers in various international journals and peer-​reviewed
conferences. He has authored and edited more than 22 books on intelligent
decision-​making, supply chain management, optimization techniques, and
risk and sustainability modeling. He has received numerous awards including
Best Track Paper Award, Outstanding Reviewer Award, Best Paper Award,
Outstanding Researcher Award, and university gold medal. Chatterjee is the
Editor-​in-​Chief of the Journal of Decision Analytics and Intelligent Computing. He
has also been the guest editor of several special issues in different SCIE /​Scopus
/​ESCI (Clarivate Analytics) indexed journals. He is also the Lead Series Editor
of “Smart and Intelligent Computing in Engineering” (Chapman and Hall /​
CRC Press), Founder and Lead Series Editor of “Concise Introductions to
AI and Data Science” (Scrivener–​Wiley); AAP Research Notes on Optimization
and Decision Making Theories; Frontiers of Mechanical and Industrial Engineering
(Apple Academic Press, co-​published with CRC Press, Taylor and Francis
Group), and “River Publishers Series in Industrial Manufacturing and
Systems Engineering.” Chatterjee is one of the developers of two multiple-​
criteria decision-​making methods called Measurement of Alternatives and
Ranking according to Compromise Solution (MARCOS) and Ranking of
Alternatives through Functional mapping of criterion sub-​intervals into a
Single Interval (RAFSI).

Morteza Yazdani currently works at Universidad Autónoma de Madrid,


Spain, as an assistant professor. He also worked at ESIC University and
Universidad Loyola Andalucía, Seville, for several years. Yazdani’s major
researches focus on management and operations, specifically on decision-​
making theories, supply chain management, and sustainable develop-
ment. Yazdani is a member of the editorial board of International Journal of
Productivity and Performance Management. He teaches operations management
and statistics, among other subjects. Yazdani has published in journals like
International Journal of Production Research, Expert Systems with Applications,
Computers and Industrial Engineering, and Annals of Operations Research.

Javier Pérez Rodríguez is a computer science engineer (University of


Córdoba), with a master’s degree in soft computing and intelligent systems
(University of Granada) and doctorate in information and communica-
tion technologies from the University of Granada. His area of research and
teaching are in two areas: computer science and artificial intelligence and
bioinformatics. Pérez has published 13 articles in the best journals in both

xiii
newgenprepdf

xiv The Editors

areas, all of them with a high impact index (Q1). His teaching experience
has been mainly in different and diverse undergraduate subjects within the
official teaching load assigned to the Department of Quantitative Methods of
Loyola Andalucía University.

Francisco Fernández-Navarro is an associate professor at the Universidad


Loyola Andalucía and a member of IEEE Computer Society. Formerly, he
was a research fellow, Innovation Dynamics and Computational Economics,
Advanced Concepts Team at the European Space Agency in the Netherlands.
He got his PhD in artificial intelligence and software engineering from the
University of Malaga. He is teaching forecasting and machine learning,
mathematics, and statistics, among others. Fernández Navarro has published
in top journals like Expert Systems with Applications, Applied Soft Computing,
and Neural Computing and Applications. He has supervised several PhD theses
as well.
newgenprepdf

1
Machine Learning for Smart Health Care

Rehab A. Rayan
Department of Epidemiology, High Institute of Public Health,
Alexandria University, Egypt
[email protected]

CONTENTS
1.1 Introduction..................................................................................................... 1
1.2 Major Applications of ML in Health Care.................................................. 3
1.2.1 Medical Diagnostics........................................................................... 3
1.2.2 Precision Health.................................................................................. 5
1.2.3 Monitoring Health.............................................................................. 7
1.3 Opportunities and Limitations..................................................................... 9
1.4 Conclusions................................................................................................... 10

1.1 Introduction
Machine learning (ML) incorporates advanced algorithms working on dis-
parate big data to reveal valuable trends that would be challenging to be
figured out by even skilled experts. Nowadays, ML applications spawn
several disciplines such as gaming (Silver et al. 2018), suggesting products
(Batmaz et al. 2019), and self-​driven cars (Bojarski et al. 2016) while in
medicine, the examples are the human genome project (Venter et al. 2001)
and the cancer omics (genomics and proteomics) (Zhang et al. 2019; Ellis
et al. 2013). Gathering and exploring health-​related big data has the poten-
tial to shift medicine into a data-​centered and outcomes-​focused field with
advances in detecting, diagnosing, and managing diseases. Molecular and
phenotypic datasets were developed, which covered genetic examination for
individualized cancer therapy, high-​resolution three-​dimensional anatomical
body parts’ images, histological examination of biopsies, and smartwatches
with biosensors for monitoring heart rates and alerting about abnormalities

DOI: 10.1201/9781003104858-1 1
2 Machine Learning Algorithms and Applications in Engineering

(Shilo et al. 2020). Such big data supply the raw material for a future of early,
precise diagnosis, customized therapies, and continuous monitoring of
health.
ML could promote health care by releasing the potential of health big
data. Prior applications of ML in diagnosis and care were promising in
detecting breast cancer from X-​rays (McKinney et al. 2020), finding novel
antibiotics (Stokes et al. 2020), expecting gestational diabetes early from
digital medical records (Artzi et al. 2020), and determining patients who
share a molecular signature of therapeutic responses (Zitnik et al. 2019). ML
digital pattern recognition could tackle complicated health big data in cases
where a manual exploration is infeasible and ineffective. Several diseases
include complicated modifications, which are experienced robustly and dif-
ferently among patients and need diligent exploration and cautious evalu-
ation of disparate big data to ascertain unique patterns so as to diagnose
and manage, hence, helping health care providers and researchers to find
and describe valuable insights from such big data (Rajkomar et al. 2019).
The functionality of a new algorithm could be strictly evaluated with previ-
ously proven associations between either quantitative biomarkers or quali-
tative data (which vary based on demographics and ecological exposures)
and patient health outcomes.
More research models have been developed for gathering and assembling
big data that relate attributes to a health condition, which could be applied
in training and testing ML techniques. Cancer models could accumulate
molecular profiles from testing frameworks or patients’ samples with data
on diagnosis, treatment, and prognosis; for instance, the Cancer Dependency
Map has gathered data on genomic stability, multimodal molecular profiles,
and therapeutic responses from thousands of cancer cell lines. By applying
innovative algorithms, these models could bring about a shift in know-
ledge about illnesses and advance anticipating health outcomes (Tsherniak
et al. 2017).
ML is derived from artificial intelligence (AI), which involves techniques to
facilitate machines showing learning and reasoning similar to humans. ML
stresses on building algorithms to learn from data. Major ML classes involve
supervised learning where datasets are linked to a certain outcome; categor-
ical values use classification techniques such as “healthy” or “diseased”
while continuous ones are applied in regression models such as responding
to treatment levels; semi-​supervised or unsupervised techniques to classify
data into specific sets, which could be manually tagged and linked to an out-
come; ensemble learning, in which findings from many digital frameworks
are integrated to give a final recommendation enabling more precise
estimations via facilitating frameworks that enhance scaling-​up to novel
data; deep learning (DL) that applies artificial neural networks, inspired by
similar networks in the human brain, to identify trends and relations in the
data where it is valuable operating on unstructured data like text, speech, or
images; and Bayesian learning where previous information is encrypted into
Machine Learning for Smart Health Care 3

the procedure of learning, particularly while working on poor data (Goecks


et al. 2020).
Several health datasets are big, having many dimensions. Reducing the
dimensions of datasets could enhance the functionality of ML techniques
by choosing a subgroup of proper data variables or merging variables into
a lower number, obtaining data variance, visualizing data, or predicting
models. In ML, techniques for federated learning could gradually learn data
that are spread over many locations and cannot be merged into one dataset
(Yang et al. 2019). Federated learning is crucial in several medical applications
working with protected or sensitive health data. Lately, accessing structured
health data and technical advances in data analytics tools contributed to the
growth of these techniques (Camacho et al. 2018). The ML techniques are
selected according to the accessed data along with the applications they are
intended for.
With the growing powerful ML supercomputers and the infrastructure
for gathering and analyzing data precisely, ML techniques would play key
roles in managing health such as detecting diseases precisely and early,
improving diagnosis, and better therapies. The quality of the learned data
could be advanced greatly through adopting standardized and large-​scale
electronic health records (EHRs) developed particularly for ML. Home data
collection via smartphones, home-​assisting devices such as Google Home or
Amazon Echo, and other digital devices would contribute to more advanced
ML in health. Figure 1.1 depicts how sophisticated algorithms are applied
on the health big data for smart outcomes-​based individualized health care.
Whereas in a house, ML could early detect disease, monitor compliance, and
response to therapy, in a hospital, ML could enable health care providers
to diagnose and personalize care for a freely moving patient between the
different settings.
This chapter discusses how ML could promote health care via better med-
ical diagnostics, precision health, and monitoring health, highlighting oppor-
tunities and successful early applications, and concluding with challenges
that hinder achieving the full potential of ML in health.

1.2 Major Applications of ML in Health Care


1.2.1 Medical Diagnostics
Today, medical examinations yield more data than ever thanks advances
in medical care. High-​fidelity imaging scans generate huge two-​, three-​, or
four-​dimensional images for tissues and organs, and molecular examinations
could analyze hundreds to thousands of genes or proteins. ML could digit-
ally analyze the diagnostic characteristics of such data to predict disease or
4 Machine Learning Algorithms and Applications in Engineering

FIGURE 1.1
ML applications in health.

response to therapy. Deep learning (DL) could analyze and interpret medical
images where many late pieces of research indicated that programs applying
ML such as computer-​aided detection (CAD) could interpret radiological
scans comparable to radiologists. For instance, a DL-​based CAD program
could precisely detect diabetic retinopathy (Gulshan et al. 2016) and deter-
mine all grades in situ or invasive breast cancer comparable to medical
experts (McKinney et al. 2020). Therefore, with the help of big data, DL-​based
techniques could function similarly to professionals over a spectrum of diag-
nosing jobs in medical imaging with high precision (Liu et al. 2019).
Molecular assays could determine abnormalities in genes and measure
stages of gene expression and abundant proteins in different samples such
as tissues, saliva, or blood where ML could identify complicated biomarkers
linked to different diseases, informing patients’ outcomes, and determining
proper managing planes. In cancer biology, for instance, applying positioning
nucleosome (Heitzer et al. 2019) and methylating DNA (Kang et al. 2017)
Machine Learning for Smart Health Care 5

in the blood to estimate the source tissue of a tumor, measuring the stage
of induction in the cell pathway from biopsies or other samples (Way and
Greene 2018), applying magnetic resonance imaging to estimate genetic
characteristics of brain tumors (P. Chang et al. 2018), and integrating imaging
with omics to predict outcomes in cancer patients (Chaudhary et al. 2018;
Mobadersany et al. 2018). ML also could identify patients who are deprived
of sleep via exploring blood mRNA, indicating the adverse effect of inad-
equate sleep on health (Laing et al. 2019). With the help of coordinated data
of different sources and biomarkers, the ML models are promising to work
more precisely than current techniques that are usually confined to a low set
of biomarkers reflecting only limited insights of the complicated diseases.
Cooperative human–​machine diagnostic techniques, as shown in Figure 1.1,
are promising for taking the advantages of both humans and machines where
the health care provider could arrive at a diagnosis via combining all the
accessed data including those generated through ML platforms (Ahuja 2019).
Hence, ML would computerize regular diagnosis, labeling troublesome cases
in need for further human intervention, and offering more valuable data to
reach a diagnosis (Ardila et al. 2019). Therefore, integrating knowledge from
both health care experts and innovative algorithms would advance diag-
noses however. There is a need to evaluate the biologic utility of these func-
tionalities. Hence, prior to broad installation and implementation, it is vital
to control the transparency of ML applications via optimal goals, qualities,
measurable functionalities, and constraints of certain algorithms and their
verifying processes (Cai et al. 2019). Thus, assisting health care providers
in applying ML programs to reach correct conclusions enhances decision-​
making. ML-​ based programs could grow confidence in the health care
system, enabling more knowledge about the implicit biologic pathways in
illnesses (Ching et al. 2018).
ML, coupled with more advances in techniques for clinical investigations,
requires considering the balance between detection rates for diseases,
patient outcomes, and distinct elements influencing health and quality of
life. Applying ML techniques would elevate the rate of detecting diseases,
and differentiating severe from minor diseases would be necessary to deter-
mine subgroups of diseases, choose highly effective therapies, and eliminate
overtreatment. For ML to promote health care, thoughtfully formulating clin-
ical objectives and the related testing and validation measures are required.

1.2.2 Precision Health
In precision health, a promising ML application, health care is customized
as per the patient’s disease profile. In precision oncology, an initial model
for ML in precision health, cancer is managed according to the molecular
features of the tumor. Lately, molecular biomarkers for a subject like the levels
of gene expressions or physical transformations usually inform the choice
6 Machine Learning Algorithms and Applications in Engineering

of therapies, yet variations in epigenomic, genomic, and anatomic disease


characteristics lead to largely different individual responses (Brown et al.
2019). Besides, the hundreds of likely medications and therapeutic regimens
may not be suitable for all conditions (Kurnit et al. 2018).
Multiscale predictive frameworks are promising ML techniques in dealing
with subjective variability. For instance, single-​purpose frameworks could
predict the functioning outcomes of biologic alterations, like the effect of
gene mutations on splicing and genetic expression (Xiong et al. 2015) and
transcription element-​ binding (K. M. Chen et al. 2019). ML techniques
could also forecast therapeutic response in cancer cell lines (Y. Chang et al.
2018), translate forecasting to patient tumors (Chiu et al. 2019), and predict
response to treatment according to clinical data on outcomes (Huang et al.
2018). Upcoming innovations in modeling for precision medicine are prom-
ising for the work on several dimensions for several needs. Multiscale mod-
eling would use biological big data to analyze an organism growing through
unique aspects of time and space. Now there are digital frameworks of
human–​virus interplay (Lasso et al. 2019), cell-​to-​cell interactions between
tumor cells and immune cells, or the entire cells (Metzcar et al. 2019). Hence,
digital frameworks of organs or even the whole organism, known as digital
twins (Björnsson et al. 2019), hold the potential for future where they could
forecast the potency of various medication mixtures, which have never been
tried in combination before and model the effect of illnesses on various
organs.
However, such multifactorial frameworks would aid in precise predictions
to be directly applied in therapy, and they might reach a mid-​level where
ML techniques create a graded roster of recommended treatments guiding
skillful and qualified health care providers in decision-​making. For example,
patient-​centered laboratory frameworks could test predictions from digital
models and suggest the best-​functioning ones for therapy. Such a hybrid
technique is promising where ML frameworks could significantly limit the
likely therapeutic combinations to be regarded and determine those to be
disregarded. Adding an experimental verifying stage could lend more proof
to the potential effectiveness of the predicted treatments.
Applying ML in digitally searching and mining knowledge in patients’
datasets and publications would advance precision health (Rajkomar et al.
2019). Patient datasets, often found in the EHRs, are a wealthy resource for
data on diagnosis, therapies, and responses for vast sets of patients. Previous
attempts to apply the algorithms of natural language processing for mining
published literature (Dong et al. 2018), EHRs (Shickel et al. 2018), and med-
ical notes (Kreimeyer et al. 2017) yielded interesting information about
relations between biologic mechanisms and biomarker-​based treatment or
using structured EHRs data to forecast the onset of disease (Artzi et al. 2020).
ML could leverage such knowledge for applications in precision health via
superior techniques, which could handle EHRs and published works of
Machine Learning for Smart Health Care 7

literature’s unstructured data and metadata. However, the existing data is


incomplete or inaccurate, making the generation of meaningful relations
difficult.

1.2.3 Monitoring Health
Treating complex diseases would be shifted from curative toward man-
aging approaches. Such a wide approach for managing health would pre-
serve health across several diseases and the natural aging process. Managing
health would need continuous monitoring of all health aspects for likely
diseases and personalizing therapies according to patients’ responses and
hence the key role of ML. Figure 1.2 shows integrating data and ML for inclu-
sive, ongoing, and precision health monitoring where data collected at both
houses or clinics are combined through predictive frameworks. Inclusive
frameworks are promising for better performance since they integrate more
personal data and are adaptable in any setting.
Beyond the clinical environment, wearables and intelligent home elec-
tronics could be used to manage health by gathering vast quantities of high-​
quality health-​related data to be used by ML techniques for recommending
timely interventions, changing lifestyle, or referring to a health care provider
for consultations. Today, wearables have in-​ built biosensors to monitor
movement, pulsation, rates of respiration, and levels of oxygen, body tem-
perature, and blood pressure, among other indicators. Test models showed

FIGURE 1.2
Inclusive ML framework.
8 Machine Learning Algorithms and Applications in Engineering

that data from wearables could be used for managing diabetes (Chang et al.
2016), detecting atrial fibrillation (Bumgarner et al. 2018) or early diagnosis
of Parkinson’s disease (Lonini et al. 2018), tracking cholesterol levels in the
blood (Fu and Guo 2018), complying with therapeutic regimens (Car et al.
2017), and to prompt an alert on a cardiac arrest (Sahoo, Thakkar, and Lee
2017). Speech-​based home aids could identify agonal breathing, an audible
early sign of a heart attack (Chan et al. 2019). Soon, ML applications would be
promising in detecting more biomarkers from audio and wearable sensors’
data, maybe via integrating data among various platforms where DL and
classical supervised learning might design models from such data.
Applying ML to the gathered data via smartphones is also promising for
diagnostics. DL techniques could analyze smartphone-​captured photos to
detect various forms of dermal tumors (Esteva et al. 2017) and to identify dia-
betes retinopathy (Micheletti et al. 2016). Lately, researchers have found that
smartphone-​gathered sensory data such as voice, response time, and acceler-
ometer data could be processed applying ML to monitor symptoms and pro-
gress of Parkinson’s disease (Ginis et al. 2016). These tested models showed
that ML-​ based wearables, home devices, and smartphones could collect
valuable data involving biometric measures, images, diet consumption, and
ecological data (Vermeulen et al. 2020). Linking such data with diagnoses,
ML could recognize patterns within the data and point to a certain diagnosis.
Managing health implies continuous monitoring of the subject’s body
functions and behavior via wearables and home devices coupled with results
from regular blood examinations. Using baseline activities and functions,
individualized frameworks could be generated by tailoring population-​wide
frameworks to gathered personal data, enabling accumulating individual
baselines and detecting deviations, which might reflect changes in health.
Applying individualized frameworks, ML techniques could monitor all
abnormalities in a person and alert them if a consultation with health care
providers should be done. Besides, tracking subjects who are searching online
for their symptoms, for example, self-​reported health problems like losing
weight, bronchitis, coughing, chest pain, among others, along with machine-​
trained multipersonal patterns, could be used to detect pancreas and lung
tumors early (White and Horvitz 2017; Paparrizos, White, and Horvitz 2016),
hence notifying a health care professional or a patient suggesting seeking
medical care if a further severe condition might convey the apparently
simple searched symptoms. However, several privacy-​related problems are
expected.
Using ML, health care providers would be able to interpret the highly pre-
cise molecular testing and imaging to recognize significant biomarkers and
reach a final diagnosis. Digital search results and multidimensional modeling
for dissimilar patients would enable diagnosing diseases requiring therapies
and provide informed therapeutic options. Following diagnosis and therapy,
managing health starts again with continuous monitoring of personal health.
Machine Learning for Smart Health Care 9

Yet, an ML framework should monitor personal therapeutic responses,


observe any side effects, and monitor general health or unusual deviations
from normal. ML would assist in implementing the first individualized frame-
work involving data on diagnosis and treatment, generating an estimated
direction for therapy regarded as the new normal.
ML could manage personal health over a lifespan via integrating and
modeling data digitally and smartly. Managing health includes creating
individualized frameworks for monitoring and precisely discovering abnor-
malities, supporting health care providers in digital diagnosis and treatment
using patients’ data, and upgrading personal frameworks with recent diag-
noses and therapies. Thus, it is data-​intensive process that needs digital
pattern for recognizing complex datasets. Managing health would be an
ongoing learning process since frameworks would be upgraded with fresh
data. Therefore, either creating novel forecasting frameworks or upgrading
available ones along with further effort are required for identifying the
potentials and barriers of such techniques with various applications.

1.3 Opportunities and Limitations


ML could contribute to transformative approaches in diagnosis and therapy;
however, it is required to grow well-​structured and high-​quality datasets for
enhancing the predictive power of ML techniques and limiting the amount of
the required data for training and the complexity of the output representations.
Speeding up in ML techniques for image recognition happened during the
introduction of ImageNet, a group of tagged and ontologically attached
images, and likewise efforts in biomedicine are required (Deng et al. 2009).
High-​quality datasets for using ML in diagnosis and therapy needs
tackling financial, technical, and legal obstacles, which usually lead to
scattered and unstandardized data. Federated learning could technically
solve this issue via integrating scattered data across systems since actual
moving of data is not required and personal privacy could be safeguarded.
Home devices and wearables could be used to gather precise data, and
ML could process such data to extract precise medical and analytic infor-
mation from unstructured origins such as published literature and EHRs.
Developing regulations for securely managing and analyzing private
health information (PHI), and legal and social standards, which deter-
mine adherence with them should be created. Both biomedical subjects
and entities should be rewarded to participate in standardizing and
sharing data. Likewise, drug and insurance companies supporting bio-
medical research should invest in the infrastructure along with collecting
and analyzing data toward better quality datasets.
10 Machine Learning Algorithms and Applications in Engineering

Rewarding techniques for sharing data, which boost variable datasets for
learning, are also required, such as local and global standards for sharing
data, which acquire data from both large hospitals and small clinics. ML
applications, which enhance patients’ therapeutic responses in large hospitals,
might not suit small clinics because of variability in patients and general
care. Yet, the ideal target of gathering health-​related data for ML requires
acquiring data from proper representative patients’ populations for building
precise ML frameworks that can be generalized to the diverse public. Hence,
focused attempts considering factors like patient condition before therapy,
therapeutic regimes, age, sex, race, ethnicity, and ecological risks are needed
(Goecks et al. 2020).
A thorough analysis of ML techniques applied in medicine is required, par-
ticularly with ongoing learning. The functionality of an ML system is best
tested through the precision of longitudinal predictions. An iterative ML tech-
nique might involve training with retrospective data, deploying algorithms,
and evaluating precision of the acquired predictions. Deployment-​gathered
data, and further retrospective big data, could retrain and enhance the algo-
rithm, followed by a cycle of testing for deployment. Assessing ongoing
learning systems, like those used in monitoring health, which adapt to changes
in behaviors or health conditions, would possibly need strengthening such
cycle and applying deployment-​gathered data in identifying both barriers
and flaws. Besides, measuring confidence intervals is vital, as some ML
applications would be further tolerating uncertain predictions than others,
hence confidence intervals could inform making decisions.
Iteratively training and deploying applications of ML holds regulatory
limitations since the majority of diagnostic and therapeutic applications pre-
sume fixed data models. Upgrading models with novel data or adapting
them for new diagnoses and therapies would require continuous assessment
for the reliability and precision of the predictions. Thus, there is a need for
real or simulated datasets, which are prospective and multidimensional and
also costly for robustly assessing the applications of ML in medicine.

1.4 Conclusions
ML incorporates advanced algorithms working on disparate big data to
reveal valuable trends that would be challenging to be figured out by even
skilled experts. ML could promote health care by releasing the potential of
health big data. This chapter discussed how ML could promote healthcare via
better medical diagnostics and precision or monitoring health, highlighting
opportunities and successful early applications, and ending with challenges
that hinder achieving the full potential of ML in health. With the growing
powerful ML supercomputers and the infrastructure for gathering and
Machine Learning for Smart Health Care 11

analyzing data precisely, ML techniques would play key roles in managing


health such as detecting diseases precisely and early, improving diagnosis,
and better therapies. The quality of the learned data could be advanced
greatly by adopting standardized and large-​scale electronic health records
(EHRs) developed particularly for ML. Despite considerable limitations, this
problem could be addressed. More research models have been developed
for gathering and assembling big data that relate attributes to a health con-
dition, which could be applied in training and testing ML techniques. By
applying innovative algorithms, these models could bring about a shift in
knowledge about illnesses and advance anticipating health outcomes. More
efforts are needed for a promising future of rigorous, outcomes-​centered
health care with detecting, diagnosing, and therapeutic protocols adjusted
through ML to personal and ecological diversities and allowing inclusive
management of health. ML would assist in the implementation of the first
individualized framework involving data on diagnosis and treatment, gener-
ating an estimated direction for therapy regarded as the new normal.

References
Ahuja, Abhimanyu S. 2019. “The Impact of Artificial Intelligence in Medicine on the
Future Role of the Physician.” PeerJ 7 (October). doi:10.7717/​peerj.7702
Ardila, Diego, Atilla P. Kiraly, Sujeeth Bharadwaj, Bokyung Choi, Joshua J. Reicher, Lily
Peng, Daniel Tse, et al. 2019. “End-​to-​End Lung Cancer Screening with Three-​
Dimensional Deep Learning on Low-​ Dose Chest Computed Tomography.”
Nature Medicine 25(6): 954–​61. doi:10.1038/​s41591-​019-​0447-​x
Artzi, Nitzan Shalom, Smadar Shilo, Eran Hadar, Hagai Rossman, Shiri Barbash-​
Hazan, Avi Ben-​Haroush, Ran D. Balicer, Becca Feldman, Arnon Wiznitzer, and
Eran Segal. 2020. “Prediction of Gestational Diabetes Based on Nationwide
Electronic Health Records.” Nature Medicine 26 (1): 71–​ 76. doi:10.1038/​
s41591-​019-​0724-​8
Batmaz, Zeynep, Ali Yurekli, Alper Bilge, and Cihan Kaleli. 2019. “A Review on Deep
Learning for Recommender Systems: Challenges and Remedies.” Artificial
Intelligence Review 52(1): 1–​37. doi:10.1007/​s10462-​018-​9654-​y
Björnsson, Bergthor, Carl Borrebaeck, Nils Elander, Thomas Gasslander, Danuta
R. Gawel, Mika Gustafsson, Rebecka Jörnsten, et al. 2019. “Digital Twins to
Personalize Medicine.” Genome Medicine 12(1): 4. doi:10.1186/​s13073-​019-​0701-​3
Bojarski, Mariusz, Davide Del Testa, Daniel Dworakowski, Bernhard Firner, Beat
Flepp, Prasoon Goyal, Lawrence D. Jackel, et al. 2016. “End to End Learning
for Self-​Driving Cars.” ArXiv:1604.07316 [Cs], April. http://​arxiv.org/​abs/​
1604.07316.
Brown, Benjamin P., Yun-​ Kai Zhang, David Westover, Yingjun Yan, Huan Qiao,
Vincent Huang, Zhenfang Du, et al. 2019. “On-​Target Resistance to the Mutant-​
Selective EGFR Inhibitor Osimertinib Can Develop in an Allele Specific Manner
Dependent on the Original EGFR Activating Mutation.” Clinical Cancer Research,
12 Machine Learning Algorithms and Applications in Engineering

January. American Association for Cancer Research. doi:10.1158/​1078-​0432.


CCR-​18-​3829
Bumgarner, Joseph M., Cameron T. Lambert, Ayman A. Hussein, Daniel J. Cantillon,
Bryan Baranowski, Kathy Wolski, Bruce D. Lindsay, Oussama M. Wazni, and
Khaldoun G. Tarakji. 2018. “Smartwatch Algorithm for Automated Detection
of Atrial Fibrillation.” Journal of the American College of Cardiology 71 (21): 2381–​
2388. doi:10.1016/​j.jacc.2018.03.003
Cai, Carrie J., Samantha Winter, David Steiner, Lauren Wilcox, and Michael Terry.
2019. “ ‘Hello AI’: Uncovering the Onboarding Needs of Medical Practitioners
for Human-​ AI Collaborative Decision-​ Making.” Proceedings of the ACM on
Human–​Computer Interaction 3 (CSCW): 104:1–​104:24. doi:10.1145/​3359206
Camacho, Diogo M., Katherine M. Collins, Rani K. Powers, James C. Costello, and
James J. Collins. 2018. “Next-​ Generation Machine Learning for Biological
Networks.” Cell 173(7): 1581–​1592. doi:10.1016/​j.cell.2018.05.015
Car, Josip, Woan Shin Tan, Zhilian Huang, Peter Sloot, and Bryony Dean Franklin. 2017.
“EHealth in the Future of Medications Management: Personalisation, Monitoring
and Adherence.” BMC Medicine 15 (April). doi:10.1186/​s12916-​017-​0838-​0
Chan, Justin, Thomas Rea, Shyamnath Gollakota, and Jacob E. Sunshine. 2019.
“Contactless Cardiac Arrest Detection Using Smart Devices.” NPJ Digital
Medicine 2: 52. doi:10.1038/​s41746-​019-​0128-​7
Chang, P., J. Grinband, B. D. Weinberg, M. Bardis, M. Khy, G. Cadena, M.-​Y. Su,
et al. 2018. “Deep-​Learning Convolutional Neural Networks Accurately Classify
Genetic Mutations in Gliomas.” AJNR. American Journal of Neuroradiology 39
(7): 1201–​1207. doi:10.3174/​ajnr.A5667
Chang, Shih-​Hao, Rui-​Dong Chiang, Shih-​Jung Wu, and Wei-​Ting Chang. 2016. “A
Context-​Aware, Interactive M-​Health System for Diabetics.” IT Professional 18
(3): 14–​22. doi:10.1109/​MITP.2016.48
Chaudhary, Kumardeep, Olivier B. Poirion, Liangqun Lu, and Lana X. Garmire. 2018.
“Deep Learning–​Based Multi-​Omics Integration Robustly Predicts Survival in
Liver Cancer.” Clinical Cancer Research 24(6): 1248–​1259. doi:10.1158/​1078-​0432.
CCR-​17-​0853
Chen, Kathleen M., Evan M. Cofer, Jian Zhou, and Olga G. Troyanskaya. 2019.
“Selene: A PyTorch-​Based Deep Learning Library for Sequence Data.” Nature
Methods 16(4): 315–​318. doi:10.1038/​s41592-​019-​0360-​8
Ching, Travers, Daniel S. Himmelstein, Brett K. Beaulieu-​Jones, Alexandr A. Kalinin,
Brian T. Do, Gregory P. Way, Enrico Ferrero, et al. 2018. “Opportunities and
Obstacles for Deep Learning in Biology and Medicine.” Journal of the Royal
Society Interface 15 (141): 20170387. doi:10.1098/​rsif.2017.0387
Chiu, Yu-​Chiao, Hung-​I Harry Chen, Tinghe Zhang, Songyao Zhang, Aparna Gorthi,
Li-​Ju Wang, Yufei Huang, and Yidong Chen. 2019. “Predicting Drug Response
of Tumors from Integrated Genomic Profiles by Deep Neural Networks.” BMC
Medical Genomics 12(1): 18. doi:10.1186/​s12920-​018-​0460-​9
Deng, Jia, Wei Dong, Richard Socher, Li-​Jia Li, Kai Li, and Li Fei-​Fei. 2009. “ImageNet: A
Large-​Scale Hierarchical Image Database.” In 2009 IEEE Conference on Computer
Vision and Pattern Recognition, 248–​255. doi:10.1109/​CVPR.2009.5206848
Dong, Wei, Xiaoling Wang, Zhi Xia, Xiuqing Zhang, and Huanming Yang. 2018. “A
Legacy of the ‘1% Program’—​The ‘Chinese Chapter’ of the Human Genome
Reference Sequence.” Journal of Genetics and Genomics—​ Yi Chuan Xue Bao
45(11): 565–​568. doi:10.1016/​j.jgg.2018.10.003
Machine Learning for Smart Health Care 13

Ellis, Matthew J., Michael Gillette, Steven A. Carr, Amanda G. Paulovich, Richard D.
Smith, Karin K. Rodland, R. Reid Townsend, et al. 2013. “Connecting Genomic
Alterations to Cancer Biology with Proteomics: The NCI Clinical Proteomic
Tumor Analysis Consortium.” Cancer Discovery 3 (10): 1108–​1112. doi:10.1158/​
2159-​8290.CD-​13-​0219
Esteva, Andre, Brett Kuprel, Roberto A. Novoa, Justin Ko, Susan M. Swetter, Helen
M. Blau, and Sebastian Thrun. 2017. “Dermatologist-​Level Classification of Skin
Cancer with Deep Neural Networks.” Nature 542 (7639): 115–​118. doi:10.1038/​
nature21056
Fu, Yusheng and Jinhong Guo. 2018. “Blood Cholesterol Monitoring with Smartphone
as Miniaturized Electrochemical Analyzer for Cardiovascular Disease
Prevention.” IEEE Transactions on Biomedical Circuits and Systems 12(4): 784–​790.
doi:10.1109/​TBCAS.2018.2845856
Ginis, Pieter, Alice Nieuwboer, Moran Dorfman, Alberto Ferrari, Eran Gazit,
Colleen G. Canning, Laura Rocchi, Lorenzo Chiari, Jeffrey M. Hausdorff, and
Anat Mirelman. 2016. “Feasibility and Effects of Home-​Based Smartphone-​
Delivered Automated Feedback Training for Gait in People with Parkinson’s
Disease: A Pilot Randomized Controlled Trial.” Parkinsonism & Related Disorders
22 (January): 28–​34. doi:10.1016/​j.parkreldis.2015.11.004
Goecks, Jeremy, Vahid Jalili, Laura M. Heiser, and Joe W. Gray. 2020. “How Machine
Learning Will Transform Biomedicine.” Cell 181 (1): 92–​ 101. doi:10.1016/​
j.cell.2020.03.022
Gulshan, Varun, Lily Peng, Marc Coram, Martin C. Stumpe, Derek Wu, Arunachalam
Narayanaswamy, Subhashini Venugopalan, et al. 2016. “Development and
Validation of a Deep Learning Algorithm for Detection of Diabetic Retinopathy
in Retinal Fundus Photographs.” JAMA 316 (22): 2402–​ 2410. doi:10.1001/​
jama.2016.17216
Heitzer, Ellen, Imran S. Haque, Charles E. S. Roberts, and Michael R. Speicher.
2019. “Current and Future Perspectives of Liquid Biopsies in Genomics-​
Driven Oncology.” Nature Reviews Genetics 20 (2): 71–​ 88. doi:10.1038/​
s41576-​018-​0071-​5
Huang, Cai, Evan A. Clayton, Lilya V. Matyunina, L. DeEtte McDonald, Benedict B.
Benigno, Fredrik Vannberg, and John F. McDonald. 2018. “Machine Learning
Predicts Individual Cancer Patient Responses to Therapeutic Drugs with High
Accuracy.” Scientific Reports 8 (1): 16444. doi:10.1038/​s41598-​018-​34753-​5
Kang, Shuli, Qingjiao Li, Quan Chen, Yonggang Zhou, Stacy Park, Gina Lee, Brandon
Grimes, et al. 2017. “CancerLocator: Non-​Invasive Cancer Diagnosis and Tissue-​
of-​Origin Prediction Using Methylation Profiles of Cell-​Free DNA.” Genome
Biology 18(1): 53. doi:10.1186/​s13059-​017-​1191-​5
Kreimeyer, Kory, Matthew Foster, Abhishek Pandey, Nina Arya, Gwendolyn Halford,
Sandra F. Jones, Richard Forshee, Mark Walderhaug, and Taxiarchis Botsis.
2017. “Natural Language Processing Systems for Capturing and Standardizing
Unstructured Clinical Information: A Systematic Review.” Journal of Biomedical
Informatics 73: 14–​29. doi:10.1016/​j.jbi.2017.07.012
Kurnit, Katherine C., Ecaterina E. Ileana Dumbrava, Beate Litzenburger, Yekaterina
B. Khotskaya, Amber M. Johnson, Timothy A. Yap, Jordi Rodon, et al. 2018.
“Precision Oncology Decision Support: Current Approaches and Strategies for
the Future.” Clinical Cancer Research: An Official Journal of the American Association
for Cancer Research 24(12): 2719–​2731. doi:10.1158/​1078-​0432.CCR-​17-​2494
14 Machine Learning Algorithms and Applications in Engineering

Laing, Emma E., Carla S. Möller-​Levet, Derk-​Jan Dijk, and Simon N. Archer. 2019.
“Identifying and Validating Blood MRNA Biomarkers for Acute and Chronic
Insufficient Sleep in Humans: A Machine Learning Approach.” Sleep 42 (1).
doi:10.1093/​sleep/​zsy186
Lasso, Gorka, Sandra V. Mayer, Evandro R. Winkelmann, Tim Chu, Oliver Elliot, Juan
Angel Patino-​Galindo, Kernyu Park, Raul Rabadan, Barry Honig, and Sagi D.
Shapira. 2019. “A Structure-​Informed Atlas of Human-​Virus Interactions.” Cell
178(6): 1526–​1541.e16. doi:10.1016/​j.cell.2019.08.005
Liu, Xiaoxuan, Livia Faes, Aditya U. Kale, Siegfried K. Wagner, Dun Jack Fu, Alice
Bruynseels, Thushika Mahendiran, et al. 2019. “A Comparison of Deep Learning
Performance against Health-​ Care Professionals in Detecting Diseases from
Medical Imaging: A Systematic Review and Meta-​Analysis.” The Lancet Digital
Health 1(6): e271–​e297. doi:10.1016/​S2589-​7500(19)30123-​2
Lonini, Luca., Dai, Andrew., Shawen, Nicholas. et al. Wearable sensors for Parkinson’s
disease: which data are worth collecting for training symptom detection
models. npj Digital Med 1, 64 (2018). https://​doi.org/​10.1038/​s41​746-​018-​0071-​z
McKinney, Scott Mayer, Marcin Sieniek, Varun Godbole, Jonathan Godwin, Natasha
Antropova, Hutan Ashrafian, Trevor Back, et al. 2020. “International Evaluation
of an AI System for Breast Cancer Screening.” Nature 577 (7788): 89–​ 94.
doi:10.1038/​s41586-​019-​1799-​6
Metzcar, John, Yafei Wang, Randy Heiland, and Paul Macklin. 2019. “A Review of
Cell-​Based Computational Modeling in Cancer Biology.” JCO Clinical Cancer
Informatics 3: 1–​13. doi:10.1200/​CCI.18.00069
Micheletti, J. Morgan, Andrew M. Hendrick, Farah N. Khan, David C. Ziemer, and
Francisco J. Pasquel. 2016. “Current and Next Generation Portable Screening
Devices for Diabetic Retinopathy.” Journal of Diabetes Science and Technology
10(2): 295–​300. doi:10.1177/​1932296816629158
Mobadersany, Pooya, Safoora Yousefi, Mohamed Amgad, David A. Gutman, Jill S.
Barnholtz-​Sloan, José E. Velázquez Vega, Daniel J. Brat, and Lee A. D. Cooper.
2018. “Predicting Cancer Outcomes from Histology and Genomics Using
Convolutional Networks.” Proceedings of the National Academy of Sciences
115(13): E2970–​E2979. doi:10.1073/​pnas.1717139115
Paparrizos, John, Ryen W. White, and Eric Horvitz. 2016. “Screening for Pancreatic
Adenocarcinoma Using Signals from Web Search Logs: Feasibility Study and
Results.” Journal of Oncology Practice 12 (8): 737–​744. doi:10.1200/​JOP.2015.010504
Rajkomar, Alvin, Jeffrey Dean, and Isaac Kohane. 2019. “Machine Learning in
Medicine.” New England Journal of Medicine, April. Massachusetts Medical
Society. www.nejm.org/​doi/​10.1056/​NEJMra​1814​259.
Sahoo, Prasan Kumar, Hiren Kumar Thakkar, and Ming-​Yih Lee. 2017. “A Cardiac
Early Warning System with Multi Channel SCG and ECG Monitoring for Mobile
Health.” Sensors (Basel, Switzerland) 17(4). doi:10.3390/​s17040711
Shickel, Benjamin, Patrick Tighe, Azra Bihorac, and Parisa Rashidi. 2018. “Deep
EHR: A Survey of Recent Advances in Deep Learning Techniques for Electronic
Health Record (EHR) Analysis.” IEEE Journal of Biomedical and Health Informatics
22(5): 1589–​1604. doi:10.1109/​JBHI.2017.2767063
Shilo, Smadar, Hagai Rossman, and Eran Segal. 2020. “Axes of a Revolution: Challenges
and Promises of Big Data in Healthcare.” Nature Medicine 26 (1): 29–​ 38.
doi:10.1038/​s41591-​019-​0727-​5
Machine Learning for Smart Health Care 15

Silver, David, Thomas Hubert, Julian Schrittwieser, Ioannis Antonoglou, Matthew


Lai, Arthur Guez, Marc Lanctot, et al. 2018. “A General Reinforcement Learning
Algorithm That Masters Chess, Shogi, and Go through Self-​Play.” Science 362
(6419): 1140–​1144. doi:10.1126/​science.aar6404
Stokes, Jonathan M., Kevin Yang, Kyle Swanson, Wengong Jin, Andres Cubillos-​Ruiz,
Nina M. Donghia, Craig R. MacNair, et al. 2020. “A Deep Learning Approach
to Antibiotic Discovery.” Cell 180(4): 688–​702.e13. doi:10.1016/​j.cell.2020.01.021
Tsherniak, Aviad, Francisca Vazquez, Phil G. Montgomery, Barbara A. Weir, Gregory
Kryukov, Glenn S. Cowley, Stanley Gill, et al. 2017. “Defining a Cancer
Dependency Map.” Cell 170 (3): 564–​576.e16. doi:10.1016/​j.cell.2017.06.010
Venter, J. C., M. D. Adams, E. W. Myers, P. W. Li, R. J. Mural, G. G. Sutton, H. O.
Smith, et al. 2001. “The Sequence of the Human Genome.” Science (New York,
N.Y.) 291(5507): 1304–​1351. doi:10.1126/​science.1058040
Vermeulen, Roel, Emma L. Schymanski, Albert-​László Barabási, and Gary W. Miller.
2020. “The Exposome and Health: Where Chemistry Meets Biology.” Science 367
(6476): 392–​396. doi:10.1126/​science.aay3164
Way, Gregory P., and Casey S. Greene. 2018. “Discovering Pathway and Cell-​Type
Signatures in Transcriptomic Compendia with Machine Learning.” e27229v1.
PeerJ Inc. doi:10.7287/​peerj.preprints.27229v1
White, Ryen W., and Eric Horvitz. 2017. “Evaluation of the Feasibility of Screening
Patients for Early Signs of Lung Carcinoma in Web Search Logs.” JAMA Oncology
3(3): 398–​401. doi:10.1001/​jamaoncol.2016.4911
Xiong, Hui Y., Babak Alipanahi, Leo J. Lee, Hannes Bretschneider, Daniele Merico,
Ryan K. C. Yuen, Yimin Hua, et al. 2015. “RNA Splicing. The Human Splicing
Code Reveals New Insights into the Genetic Determinants of Disease.” Science
(New York, N.Y.) 347 (6218): 1254806. doi:10.1126/​science.1254806
Yang, Qiang, Yang Liu, Tianjian Chen, and Yongxin Tong. 2019. “Federated Machine
Learning: Concept and Applications.” ACM Transactions on Intelligent Systems
and Technology 10(2): 12:1–​12:19. doi:10.1145/​3298981
Zhang, Zhiqiang, Yi Zhao, Xiangke Liao, Wenqiang Shi, Kenli Li, Quan Zou, and
Shaoliang Peng. 2019. “Deep Learning in Omics: A Survey and Guideline.”
Briefings in Functional Genomics 18 (1): 41–​57. doi:10.1093/​bfgp/​ely030
Zitnik, Marinka, Francis Nguyen, Bo Wang, Jure Leskovec, Anna Goldenberg, and
Michael M. Hoffman. 2019. “Machine Learning for Integrating Data in Biology
and Medicine: Principles, Practice, and Opportunities.” Information Fusion 50
(October): 71–​91. doi:10.1016/​j.inffus.2018.09.012
newgenprepdf

2
Predictive Analysis for Flood Risk Mapping
Utilizing Machine Learning Approach

Aditya Singh,1 Sunil Khatri,2 Sandhya Save,3 and Hemant Kasturiwale4


Department of Electronics Engineering, Thakur College of Engineering & Technology,
1

Kandivali, India
2
Assistant Professor, Department of Electronics Engineering, Thakur College of
Engineering & Technology, Kandivali, India
3
Professor, Department of Electronics Engineering, Thakur College of Engineering &
Technology, Kandivali, India
4
Associate Professor, Department of Electronics Engineering, Thakur College of
Engineering & Technology, Kandivali, India

CONTENTS
2.1 Introduction................................................................................................... 18
2.1.1 Machine Learning (ML)................................................................... 20
2.1.2 ML Algorithms Used for Flood Risk Mapping............................ 20
2.1.2.1 Artificial Neural Networks............................................... 20
2.1.2.2 Multilayer Perception........................................................ 20
2.1.2.3 Adaptive Neuro Fuzzy Inference System
(ANFIS)................................................................................ 21
2.1.2.4 Wavelet Neuro Networks (WNN)................................... 21
2.1.2.5 Support Vector Machine (SVM)....................................... 21
2.2 Methodology................................................................................................. 21
2.2.1 Study Area......................................................................................... 21
2.2.2 Remote Sensing................................................................................. 24
2.2.3 LULC Map Creation......................................................................... 24
2.2.4 State Selection................................................................................... 24
2.2.5 Markov Model.................................................................................. 25
2.2.6 Transition Matrix Calculations....................................................... 25
2.3 Results............................................................................................................ 27
2.4 Conclusion..................................................................................................... 28

DOI: 10.1201/9781003104858-2 17
18 Machine Learning Algorithms and Applications in Engineering

2.1 Introduction
Land cover refers to the surface cover of the ground, whether it is vegetation,
water, or bare soil. In short, land cover indicates the physical type of the land
(water, snow, grassland, soil). Land use refers to human activities that are
directly related to the land (agricultural land, canal, built-​up land, and other
human-​made characteristics). Together, they form a pattern called as land use
and land cover (LULC) pattern. This pattern is an outcome of socioeconomic
and natural factors and their utilization by humans in time and space [1].
Land use and land cover mapping is carried out to study the land utilization
pattern and future planning and management of the land resource to either
avoid economic loss due to natural factors or improve the ecological balance in
the system. Also, it is used for planning and development of land parcel. The
assessment of risks and the development of risk maps for future land use and
infrastructure development is essential. Therefore, the change in land use land
cover (LULC) pattern is detected. This pattern is affected by floods that occur
in a particular region. Floods can cause devastation to human lives, property,
and possessions as well as disruptions in communications. Flood tends to occur
when rainfall is very high, absorption is very low, and when overflows are not
controllable. Some other factors responsible for floods are unconditional rainfall,
increase in the number of low-​lying areas, rising sea level, poor sewage system,
and low absorption capability of soil. Using predictive analysis and remote
sensing methods, floods can be predicted in advance, which will not only help
in detecting imminent patterns but will also help us in preparing beforehand
to tackle such situations and come up with proper contingency plans. We plan
to model and predict future impacts as well as the rate of occurrence of floods
using image processing and predictive analysis.
Earlier, when no remote sensed data and computer assistance were avail-
able, land use/​land cover changes were detected with the aid of tracing
papers and topographical sheets. But this method was tedious, inefficient,
and inaccurate. Studying large areas required considerable amount of effort
and time. Conventional ground methods used for land use mapping are
labor-​intensive, time-​consuming, and are done less frequently [2]. Thus,
with the advent of satellite remote sensing techniques, which allow for easy
collection of data over a variety of scales and resolutions, preparing accurate
land use land cover maps was feasible. Also, monitoring changes at regular
intervals of time became relatively simpler. In case of an inaccessible region,
the only method of acquiring required data is by applying this technique.
Today, remote sensing and geographical information system (GIS) tech-
nology has enabled ecologists and natural resources managers to acquire
timely data and observe periodical changes. Predicting floods in any location
has remained a longstanding challenge, which plays a major role in emer-
gency management.
Predictive Analysis for Flood Risk Mapping 19

Remote sensing makes it easier to locate floods that have spread over a
large region, thereby making it easier to plan rescue missions quickly. Also,
remote sensing is a relatively cheap and efficient method in comparison to
traditional methods. Predictive analytics offers a unique opportunity to iden-
tify certain trends and patterns that are used to identify future outcomes.
Implementation of predictive mapping techniques became easier with the
advent of predictive modeling techniques and their simple incorporation
with other technologies. Thus, by combining both remote sensing and pre-
dictive analysis, a model can be created that will help in the detection of
floods well in advance.
High-​definition satellite images can be integrated with socioeconomic
data in order to build environmental, economic, and social threat prediction
models. Using these models, one can build demand-​driven applications to
help public and private organizations understand, prepare, and respond to
economic, social, and humanitarian losses in a timely manner. Such models
can be used to make near-​real-​time remote sensing applications by adding
additional layers of response tools, like alerts and advisories. Ideally, these
technologies would be able to be scaled up from local to global for emergency
management and risk management. Some of the benefits it can offer are dis-
aster risk reduction, event prediction for timely response and recovery, and
allowing stakeholders to better target investments and protect assets. With
the help of this study, governments can formulate suitable policies that are in
the interest of the people and which will also help us in attaining ecological
stability.
Among the cataclysmic events, floods are the most dangerous in
causing severe harm to human life, foundation, farming, and the finan-
cial frameworks [1]. Flood risk analysis is undertaken to research land use
and is also used in the planning and management of future land resources,
which is done either to prevent economic loss due to natural causes or to
enhance the system’s ecological balance. Using predictive analysis and
remote sensing methods, floods can be predicted in advance, which will not
only help in detecting imminent patterns but will also help us in preparing
beforehand to tackle such situations and come up with proper contingency
plans. Implementation of predictive mapping techniques became simpler
with the advent of predictive modeling techniques and its simple incorpor-
ation with other technologies [1]. Robust and precise prediction significantly
leads to water resource management strategies, policy recommendations
and research, as well as further forecasting and evacuation. Governments
are actually under pressure to set up dependable and precise maps of flood
risk regions and further prepare for effective management of floods [1].
Governments will be able to devise suitable policies with the help of this
study that are in people’s interest and which will eventually help to achieve
ecological stability.
20 Machine Learning Algorithms and Applications in Engineering

2.1.1 Machine Learning (ML)


A data-​driven model is a type of model that uses ML techniques to build a
relationship between input and output data, without having to worry too
much about the underlying process. ML is a part of artificial intelligence
(AI) process for inducing regularities and patterns, making implementation
simpler, with low computation costs, quick training, validation, testing and
evaluation, high performance compared to physical models, and compara-
tively less complex compared to physical models. One of the characteristics
of ML algorithms that must be considered is that these models are a good
as their training where the system, based on past data, learns the target
tasks [1].
If the data is sparse or does not cover a variety of tasks, their learning
will be limited, and therefore they will be unable to perform well when put
into work.

2.1.2 ML Algorithms Used for Flood Risk Mapping


2.1.2.1 Artificial Neural Networks
Artificial neural networks is a model of computation based on biological
neural network structure and functions. They are considered as a nonlinear
modeling tool for statistical data where the complex associations between
inputs and outputs are designed or patterns are associated [1]. As one of the
most popular learning algorithms, ANNs are known to be flexible and effi-
cient in modeling complex flood processes with large fault tolerance and
accurate approximations, among the ML methods. ANN utilizes historical
data and is more reliably used for flood prediction. ANN offers the most
appropriate modeling technique and acceptable generalization capacity and
speed compared to other models. Every ANN layer essentially carries out
nonlinear input transformations from one vector space to another. The rela-
tively subpar accuracy, the urge for iterative parameter adjustment, and the
sluggish response to gradient-​based learning process are major drawbacks to
using ANN.

2.1.2.2 Multilayer Perception
For a more sophisticated modeling approach, the knack percolation learning
algorithm calculates the propagation error in hidden network nodes indi-
vidually. Nonlinear activation functions resolve the major drawbacks of
linear activation function by allowing backpropagation because they have
an input-​related derivative function [1]. And they allow various layers of
neurons to stack to create a deep neural network. Because of many variables,
MLP is considered to be tougher to optimize.
Predictive Analysis for Flood Risk Mapping 21

2.1.2.3 Adaptive Neuro Fuzzy Inference System (ANFIS)


Fuzzy logic as a simplified mathematical model works to incorporate
the knowledge of the expert into a fluid inference system. Fuzzy interfer-
ence system (FIS) further imitates human learning through a less complex
approximation feature, which offers great potential for nonlinear designing
of extreme hydrological occurrences such as floods. Neurofuzzy has a back-​
propagation neural network (BPNN) variant and the least square test of
error, quick and easy to implement, effective learning, and good generaliza-
tion skills [1].

2.1.2.4 Wavelet Neuro Networks (WNN)


Wavelets transform (WT) can be used to analyze local time-​series variation
to extract information from different data sources. To boost data quality,
it supports accurate decomposition of an original time sequence. This
incorporates wavelet theory and the neural network (NN) in one theory.
WNN, which incorporates WT and FFNN, and the wavelet-​based regression
model, which integrates WT and multiple linear regression (LR), were used
to forecast floods in time series [1].

2.1.2.5 Support Vector Machine (SVM)


SVM is a supervised learning machine that functions on the theory of statis-
tical learning and the rule of structural risk minimization. With a huge popu-
larity among hydrologists for flood prediction, SVM and SVR were used as
the alternative ML methods to ANNs [1]. It is ideal for linear and nonlinear
classification as well as efficient input mapping into feature spaces. However,
a high cost of computation and disappointing performance make it a lesser
preferred option.
Table 2.1 presents a survey of the literature on work done by earlier
researchers on similar tracks.

2.2 Methodology
2.2.1 Study Area
Mumbai is the capital of the state of Maharashtra, one of states with the
highest population. It is located on a peninsula on the western coast of India
sharing borders with Arabian Sea to the south and west, Mira Bhayander
and Thane to the north and Navi Mumbai to the east [2]. The study area, that
is, Mumbai suburban district, has an area of 446 sq. km. and a population
22 Machine Learning Algorithms and Applications in Engineering

TABLE 2.1
Literature Review of Work Done by Various Authors over Similar Tracks

Presented at Year Name Author Inference

Journal of 2018 Remote sensing M. T. Esetlili, Prediction of


Environmental and geographic Y. Kurucu, probable
Protection and information system G. Cicek, reduction in
Ecology in the management O. Demirtas the yields
of agricultural risks that might
related to climate occur through
change monitoring and
assessment.
International Journal 2009 Analysis of urban B. Bhatta Study of urban
of Remote Sensing growth pattern using growth pattern
remote sensing and
GIS
University of Delhi, 2018 Linear regression Khushbhu Kumari, The concept of
New Delhi analysis and study Suniti Yadav correlation
and linear
regression was
understood and
the relationship
between them
was obtained.
G. Rudolph et al. 2008 Prediction using linear AnabelaSimoes, Understanding of
regression and Ernesto Costa Markov chains
Markov chains and dynamic
environment
prediction
Landscape and 2001 Predicting land-​cover Erna Lo´peza, Markov chain
Urban Planning and land-​use change Gerardo implementation
55 in the urban fringe Boccoa, Manuel for prediction of
Mendozaa, Emilio change in land
Duhaub features
Computer 2000 Link prediction and Ramesh R. Sarukkai Understanding
Networks 33 path analysis using fundamentals of
Markov chains Markov chain
algorithms
The R Journal 2017 RQGIS: Integrating JannesMuenchow, Understanding
R with QGIS for PatrikSchratz, the R package
statistical geo-​ Alexander RQGIS
computing Brenning
Journal of Geomatics 2019 Spatial pattern of Akintunde John Seeks to critically
urban growth using Akinrinola understand the
remote sensing and pattern of urban
landscape metrics growth
Sustainability 2017 Land use/​land cover Wei Wu, Eshetu Explores a
change modeling Yirsaw, Xiaoping future LULC
and the prediction of Shi, Habtamu simulation using
subsequent changes Temesgen, Belew a CA-​Markov
in ecosystem service Bekele model
values in a coastal
area of China
Predictive Analysis for Flood Risk Mapping 23

TABLE 2.1 (Continued)
Literature Review of Work Done by Various Authors over Similar Tracks
Presented at Year Name Author Inference

National Institute 2011 Land use and land Biswajit Majumdar Mapping of land
of Technology, cover change use and land
Rourkela detection at Sukinda cover pattern
Valley using remote
sensing and GIS
Journal of Applied 2018 Land use/​land cover Bello Ho Markov based
Science change analysis LULC changes
Environment using Markov-​based with respect to
Management model for Eleyele water bodies
Reservoir
International 2018 Accuracy analysis of Marina Gudelj, Workflow for this
Multidisciplinary the inland waters Mateo Gasparovic, project
Scientific Geo detection MladenZrinjski
Conference
Landscape and 2016 Specific features of Szilard Szabo, NDWI, NDVi,
Environment 10 NDVI, NDWI and Zoltan Gasci, MNDWI
MNDWI as Reflected BoglarkaBalazs coverages for
in LC Categories water bodies
LULC study

FIGURE 2.1
Geographical Region of Mumbai Suburban Region.

of 9,356,962 according to the 2011 census [3]. Thus, the density of Mumbai
suburban district is 20,980 people per square kilometer. With elevations ran-
ging from 10 meters to 15 meters, the city generally lies just above the sea
level with an average elevation of 14 meters [4, 5]. This makes the city highly
prone to flooding and waterlogging. Figure 2.1 shows a geographical map of
Mumbai.
24 Machine Learning Algorithms and Applications in Engineering

2.2.2 Remote Sensing
Earlier, when no remote sensed data was available, changes in LULC
patterns were detected using tracing papers and topographic sheets. But that
approach has been slow, unsuccessful, and unreliable. Wide areas of study
involved a considerable amount of time, effort, and cost [5, 6]. Conventional
solutions are labor-​intensive, time-​consuming, and less often performed.
Therefore, the preparation of detailed land use and land cover maps and
tracking adjustments at regular intervals of time is comparatively simpler
with the advent of satellite remote sensing techniques [7]. In the event of
inaccessible areas, the only method of data acquisition is by applying this
technique. Quantum geographical information system (QGIS) is the software
tool used for remote sensing purposes.

2.2.3 LULC Map Creation


LISS—​III images were used in the study marked in the years 2011 to 2015.
The processing of the image and classification was carried out using QGIS
version 3.3 Zanzibar [8, 9]. Normalized difference water index (NDWI) is
used to monitor changes related to water content in water bodies, using
green and NIR wavelength.

NDWI = ( NIR − SWIR ) ( NIR + SWIR )

The Bands of the LISS-​III Satellite and their Wavelengths are shown in
Table 2.2.

2.2.4 State Selection
Since the Markov model is a stochastic state-​ based transition model,
selection of states is an important step. The states need to be clearly distin-
guishable from the available LULC maps as that will provide greater
accuracy [10].

TABLE 2.2
Bands of LISS-​III Satellite and Their Respective Wavelengths

Band Wavelength Revisit time

BAND 2(VIS) 0.52 to 0.59 24


BAND 3(VIS) 0.62 to 0.68 24
BAND 4(NIR) 0.77 to 0.86 24
BAND 5(SWIR) 1.55 to 1.75 24
Predictive Analysis for Flood Risk Mapping 25

2.2.5 Markov Model
We have seen that whenever a succession of chance analyses frames a free
preliminaries process, the potential results for each examination are equiva-
lent and happen with a similar likelihood [11, 12]. Further, information on
the results of the past investigations doesn’t impact our expectations for the
results of the following test. The circulation for the results of a solitary trial
is adequate to develop a tree and a tree measure for a grouping of n trials,
and we can respond to any likelihood question about these trials by utilizing
this tree measure. Current likelihood hypothesis reads chance procedures for
which the information of past results impacts expectations for future ana-
lyses. At a fundamental level, whenever we watch a grouping of chance
analysis, the entirety of the past results could impact our forecasts for the
following test [6, 12].
Markov chain calculates how much land is estimated to change from the
latest date to the expected date [7]. In this method, the transfer probabilities
are the output, which is a matrix that records the likelihood that each land
cover class will move to each other class[9, 12]. Through the Markov chain
simulation, the analysis of two different dates of the LULC images creates
the transition matrices, a transition area matrix, and a set of conditional prob-
ability images. The transition probability matrix for the period of 2008–​2015
was determined to predict the 2019 LULC map and thus any future given
date for prediction [13]. The probability of one pixel moving to another
LULC or staying in the original LULC can be calculated by producing a tran-
sition matrix of probabilities. The transition matrix is generally important
for predicting a future classification map [14]. If the hypothesis tested has no
significant differences in the observed LULC and the predicted LULC, the
model is considered successful for future predictions [15, 16].

2.2.6 Transition Matrix Calculations


Figure 2.2 shows State Markov Model. LISS—​III images are used for remote
sensing analysis. This part includes preprocessing of images followed by

FIGURE 2.2
State Markov Model.
26 Machine Learning Algorithms and Applications in Engineering

FIGURE 2.3
Block Diagram of the proposed model.

supervised classification. NDWI is calculated in this block. After accuracy


assessment, the corresponding maps are generated [17]. The Markov model
depends and makes transitions between a fixed number of states. The
states are classified as transient and absorbing states [18]. The absorbing
states mark an end to the process and thus need to be selected accordingly.
The initial state transition probabilities are determined based on the basis
of the initial land cover of a specific state [19]. The transition matrix can be
determined by a change in the first iteration. This determines the number
of states or in real world, the time required by the process to end up in an
absorbing state [20]. This will obviously exist for all states since the given
problem is an absorbing state problem, but we will estimate the transition
to the absorption state in a finite amount of time. Multiple iterations
will be carried out and finally we will be able to see the trends or pattern
this model will follow [21, 22]. Figure 2.3 shows Block Diagram of the
proposed model.
Predictive Analysis for Flood Risk Mapping 27

FIGURE 2.4
NDWI Map of the year 2011.

2 . 3 
R esults
Using QGIS software as a tool for remote sensing, the study area was analyzed
for a period of five years starting from 2011 to 2015. During this analysis,
NDWI values were calculated for each year and their corresponding maps
were generated. Figure 2.4 shows NDWI Map of the year 2011. The NDWI
maps and rainfall data for year 2011–​2015 are shown in Figure 2.5. In order
to have a clear picture of the water bodies present in the study area, Band 5
of LISS—​III satellite is used for short-​wave infrared sensing. These NDWI
values, along with rainfall data in the above-​mentioned years, are taken
as parameters or variables for carrying out predictive analysis using the
Markov chain model. Figure 2.6 shows NDWI Map of the year 2015 whereas
Figure 2.7 shows Rainfall Data of the year 2015.
The Markov model, a stochastic state-​based transition model, is used with
two different states. The states need to be clearly distinguishable from the
available LULC maps as it will provide greater accuracy. In this case, two
states are decided. The first state is the state of being in flood (A) and the next
28 Machine Learning Algorithms and Applications in Engineering

FIGURE 2.5
Rainfall Data of the year 2011.

state is the state of not being in the flood (B). So, basically there are two states,
viz. A and B. The transition probabilities will be calculated and the transition
matrix will be formed depending upon these states. The transition matrix
calculated for years 2011–​2015 is shown in Figure 2.8.
Figure 2.9 shows the implementation of two-​state Markov model using the
MATLAB simulation tool. The time steps for simulation are 5, and the transi-
tion probabilities of each element are depicted in this simulation.

2.4 Conclusion
The geographic analysis of study area, viz. coverage area, population, and
height from sea, is performed to make a prediction about the possibility of
flooding and waterlogging. Superior quality satellite images are integrated
Predictive Analysis for Flood Risk Mapping 29

FIGURE 2.6
NDWI Map of the year 2015.

with socioeconomic data in order to build environmental, economic, and


social threat prediction models. The processing of the image and classifica-
tion is carried out using QGIS version 3.3 Zanzibar. The processing of
images is done on basis of calculation of normalized difference water NDWI
values, which are related to water content in water bodies, using green and
NIR wavelengths. Markov chains provide a stochastic model of dispersion
that applies to singular particles, which are used for stochastic dissemination
we’ve examined. This stochastic dissemination likewise gives a valuable
model of the spread of data all through the picture. This Markov chain-​based
model gives an accurate prediction of flood conditions.
Such models can be used to make near-​real-​time remote sensing applications
by adding additional layers of response tools, like alerts and advisories, for
disaster and risk management. More training data can be used to improve the
accuracy of this designed model.
30 Machine Learning Algorithms and Applications in Engineering

FIGURE 2.7
Rainfall Data of the year 2015.

FIGURE 2.8a
Two State Markov Models for the year 2011–​2015.
Predictive Analysis for Flood Risk Mapping 31

FIGURE 2.8b
Two State Markov Models for the year 2011–​2015.

FIGURE 2.9
State Markov Model Implementation on MATLAB.
32 Machine Learning Algorithms and Applications in Engineering

References
[1] Mosavi, A.; Ozturk, P.; Chau, K.-​W. Flood Prediction Using Machine Learning
Models: Literature Review. Water 2018, 10, 1536.
[2] www.cen​sus2​011.co.in/​cen​sus/​distr​ict/​357-​mum​bai-​city.html
[3] http://​pibmum​bai.gov.in/​Engl​ish/​PDF/​E201​3_​PR​798.PDF
[4] Biswajit Majumdar, “Land use and land cover change detection at Sukinda Valley
using remote sensing and GIS”, National Institute of Technology Rourkela, 2011.
[5] Rahel Hamad, Heiko Balzter, Kamal Kolo, “Predicting land use/​ land
cover changes using a CA-​ Markov model under two different scenarios”,
Sustainability, 2018.
[6] Jannes Muenchow, Patrik Schratz, Alexander Brenning, “RQGIS: Integrating R
with QGIS for statistical computing,” R Journal, 2017.
[7] S Junaida, Sulaiman and Siti Hajar, Wahab (2018) Heavy Rainfall Forecasting
Model Using Artificial Neural Network for Flood Prone Area. In: IT Convergence
and Security 2017. Lecture Notes in Electrical Engineering, 449 . Springer,
Singapore, pp. 68–​76.
[8] S. Khatri and H. Kasturiwale, "Quality assessment of Median filtering techniques
for impulse noise removal from digital images," 2016 3rd International
Conference on Advanced Computing and Communication Systems (ICACCS),
2016, pp. 1–​4, doi: 10.1109/​ICACCS.2016.7586331.
[9] J. Abbot,; J. Marohasy, “Input selection and optimization for monthly rain-
fall forecasting in Queensland, Australia, using artificial neural networks.
Atmospheric Research, pp. 166–​178, 2014.
[10] A. Jain,, S. Prasad Indurthy, Closure to “comparative analysis of event-​based
rainfall-​
runoff modeling techniques—​ deterministic, statistical, and artificial
neural networks”.. Journal of Hydrologic Engineering, vol. 9, pp. 551–​553, 2004.
[11] R.C. Deo, M. Sahin, Application of the artificial neural network model for pre-
diction of monthly standardized precipitation and evapotranspiration index
using hydrometeorological parameters and climate indices in eastern Australia.
Atmospheric Research, pp. 65–​81, 2015.
[12] M.K. Tiwari, C. Chatterjee, Development of an accurate and reliable hourly
flood forecasting model using wavelet–​bootstrap–​ANN (WBANNss) hybrid
approach. Journal of Hydrology, pp. 458–​470, 2010.
[13] C.A. Guimarães Santos, G.B.L.d. Silva. Daily streamflow forecasting using a
wavelet transform and artificial neural network hybrid models. Hydrological
Sciences Journal, pp. 312–​324, 2014.
[14] S. Supratid, T. Aribarg, S. Supharatid. An integration of stationary wavelet
transform and nonlinear autoregressive neural network with exogenous
input for baseline and future forecasting of reservoir inflow. Water Resources
Management, vol. 31, pp. 4023–​4043, 2017.
[15] E. Dubossarsky, J.H. Friedman, J.T. Ormerod,. M.P. Wand. Wavelet-​based gra-
dient boosting. Statistics and Computing, vol. 26, pp. 93–​105, 2016.
[16] K. Kasiviswanathan, J. He,; K. Sudheer, J.-​H. Tay. Potential application of wavelet
neural network ensemble to forecast streamflow for flood management. Journal
of Hydrology, pp. 161–​173, 2016.
[17] L.A. Zadeh. Soft computing and fuzzy logic. In Fuzzy Sets, Fuzzy Logic, and Fuzzy
Systems: Selected Papers by Lotfi a Zadeh, World Scientific: pp. 796–​804, 1996.
Predictive Analysis for Flood Risk Mapping 33

[18] P.S. Yu, T.C. Yang, S.Y. Chen, C.M. Kuo, H.W. Tseng.Comparison of random
forests and support vector machine for real-​ time radar-​
derived rainfall
forecasting. Journal of Hydrology, vol. 552, pp. 92–​104, 2017.
[19] S. Li, K. Ma, Z. Jin, Y. Zhu. In A new flood forecasting model based on svm
and boosting learning algorithms, Evolutionary Computation (CEC), IEEE
Conference, pp. 1343–​1348, 2016.
[20] E. Danso-​Amoako, M. Scholz, N. Kalimeris, Q. Yang, J. Shao. Predicting dam
failure risk for sustainable flood retention basins: A generic case study for the
wider greater Manchester area. Computers, Environment and Urban Systemsvol
Vol. 36, pp. 423–​433, 2012.
[21] Y.B. Dibike, S. Velickov, D. Solomatine, M.B. Abbott. Model induction with
support vector machines: Introduction and applications. Journal of Computing
in Civil Engineering, vol.15, pp. 208–​216, 2001.
[22] Rahel Hamad, Heiko Balzter, Kamal Kolo, Predicting land use/​ land cover
changes using a CA-​Markov model under two different scenarios, 2018.
newgenprepdf

3
Machine Learning for Risk Analysis

Parita Jain,1,* Puneet Kumar Aggarwal,2 Kshirja Makar,3 Riya Garg,3


Jaya Mehta,3 and Poorvi Chaudhary3
1
KIET Group of Institutions, UP, India
2
ABES Engineering College, UP, India
3
HMRITM, Delhi, India
[email protected]
*

CONTENTS
3.1 Introduction................................................................................................... 35
3.1.1 Machine Learning............................................................................. 36
3.1.2 Risk Analysis..................................................................................... 37
3.2 Risk Assessment and Risk Management................................................... 40
3.2.1 Risk Management............................................................................. 40
3.3 Risks in the Business World........................................................................ 41
3.3.1 Models in the Field of Risk Management..................................... 43
3.4 Machine Learning Techniques for Risk Management and
Assessment.................................................................................................... 44
3.4.1 Challenges of Machine Learning in Risk Management.............. 46
3.4.2 Machine Learning Use Cases in the Financial Sector.................. 48
3.5 Case Studies to Understand the Role of Machine Learning in
Risk Management......................................................................................... 49
3.6 Conclusion..................................................................................................... 51

3.1 Introduction
Machine learning (ML) is the new development in technology. It has the
proficiency of replacing emphatic programming of the devices. It is gener-
ally hinged on the conception that a substantial amount of data is provided
and on the basis of that data and some algorithms, the machine is trained

DOI: 10.1201/9781003104858-3 35
36 Machine Learning Algorithms and Applications in Engineering

and different machine modules are fabricated. The decisions made by the
machines on the basis of data provided are immensely efficient and accurate.
With increased use of technology, ingenious crimes and risks are also escal-
ating. In the modern world, the most important constituent to take off is
risk management and its intensifying production. This chapter provides all
the information regarding how machine learning has been applied in risk
assessment (Apostolakis, 2004; Aven, 2012).
Large-​scale organizations, companies, and institutions are prone to risks
like frauds; consequently, they are sticking to various machine learning
techniques that can prevent or abate these frauds or risks. The chapter is
an amalgamation of various strategies for changing the perspective of risk
assessment. On the basis of risk assessment with machine learning, various
case studies covering different aspects are presented in the chapter. These
case studies effectively portray the seriousness of risk assessment. Case
studies elucidate constructing techniques to diminish risks and act as a ref-
erence for future. Several applications are also incorporated in this chapter,
which lead us to the significance of risk assessment in several industries and
organizations for protection from a large amount of frauds and deceptions
altogether with the integration of machine learning in this peculiar field
(Chen, 2008; Cheng, 2016).
The essence of this chapter is related to the arguments that provide a solu-
tion for risks found in the literature followed by a conclusion, which covers
machine learning in the field of risk management as a whole.

3.1.1 Machine Learning
Machine learning can be described as a subcategory of Artificial Intelligence
(AI), which has its main focus on examining and recognizing patterns and
arrangements in data to facilitate features such as training, thinking, decision
making, learning, and researching without interference. Machine learning
allows the user to pack an enormous sum of data with a computer algo-
rithm and allows the computer to examine and analyze the data to make
recommendations based on the input received. If some features require
redesigning, they are classified and improved for a better design for the
future (Creedy, 2011; Comfort, 2019).
The main aim of the technology is to produce an easy-​ to-​
use mech-
anism, which works on making decisions by the computational algo-
rithm. Variables, algorithms, and innovations are accountable for making
decisions. Awareness toward the solution is needed for the better learning
of the working of the systems and thus helps in understanding the path to
reach the result.
In the initial stages of implementation of the algorithm, input or data is fed
to the machine provided the result is already known for that set of feed data.
The changes are then made to produce the result (Diekmann, 1992; Durga,
Machine Learning for Risk Analysis 37

FIGURE 3.1
Machine learning algorithm process.

2009; Goodfellow, 2016). The efficiency of the result depends on the amount
and the quantity of the data that is fed to the machine as shown in Figure 3.1.

3.1.2 Risk Analysis
Risk is referred to as the occurrence of undesired events while running a pro-
ject, which creates a negative impact on the achievement of the goals of that
project. Risk is the possibility of an unwanted or harmful event. Risk analysis
is done by adopting different methods in order to check the probability of
the happening of the risk and it is performed in order to remove the chances
or minimize the probability of the occurrence of the harmful events. Risk
analysis is desirable as the occurrence of risk is directly proportional to the
amount of losses faced in a project. Every technology or project has an equal
chance of being successful as well as becoming a failure and hence the ana-
lysis is done to reduce this chance, which is called the risk of failure (Kaplan
1981; Hastie, 2009; Hauge, 2015; Haugen, 2015; Khakzad, 2015).
Risk analysis includes identifying the type of hazard that can be associated
with the project, which may be a chemical hazard, mechanical hazard, or
even technical hazard. It is a process that is analytical in nature and the
aim is to know all the desired information related to undesirable events. It
involves an analysis of the hazards that have occurred in the past and also
which have the probability of occurring in future (Khakzad, 2013a; 2013b).
It not only analyzes the hazards but also their consequences on the system
so that appropriate measures can be taken. The main objective is to increase
the chances of success and at the same time minimize the cost or investment
on the project. As important risk analysis is, it is even more difficult to be
performed. The traditional method of using employees requires a lot of time
to complete the process (King, 2001; Kongsvik, 2015; Landucci, 2016a, 2016b).
But now with the growth of industries, the amount of data generated is very
38 Machine Learning Algorithms and Applications in Engineering

high, operations performed on a project are more than before, but the time to
complete that project is lower. So, in order to fulfill the demand versus time
ratio, industries use technical and more reliable methods such as ML. There
are different methods in which risk analysis can be done. These methods are
as follows:

(i) Qualitative methods: It is a method generally used in the process of deci-


sion making or before decision making. It involves various predictions
based on previous experiences; different judgments are passed by
different members of the project. This is not a very accurate method and
is generally done at the initial stage of the project (Jain, 2018; Jain, 2019).
The major aim is to improve the quality of the project based on past
experiences. This method is suggested when there is no time constraint
and the risk level is quite low. This method does not make use of any
algorithms or numerical data. It is performed through brainstorming,
by questionnaire, or interviews.
(ii) Quantitative methods: This method is way more numerical and reliable
than the qualitative methods. In this method, different numerical values
can be assigned to different risks identified based on the probability of
their occurrences and the algorithms are performed to calculate the level
of the risk of the project (Aggarwal 2018; Aggarwal 2019; Jain 2020). In
this method, we do not assume what is going to happen. It involves
analyzing the probability of the number of occurrences, analysis of the
consequences as well as the after-​effects of the solution. This method is
a little expensive as it requires machines and computer systems, but it
is very accurate in results.
(iii) Semiquantitative methods: These methods are neither fully qualitative nor
completely numerical or algorithm based. This is a mid-​way method
adapted according to the needs of the project and clients. The method
used depends on the level of risk of the project. In this method, various
projects are classified as low-​, middle-​, or high-​risk level projects. This
method is used in small-​scale projects or where the client is not sure
about the requirements and hence is not able to predict the exact level
of the associated risk.

The major components of risk analysis include hazard identification, risk


assessment, risk management, and risk communication (Nivolianitou, 2004;
Nobre, 2009; Musgrave, 2013; Lasi, 2014). The step-​by-​step process is shown
in Figure 3.2.

(i) Hazard identification: The primary step to proceed toward risk analysis
is the identification of the problems, risks, or hazards that can be faced
in the future. The number of hazards determines the level of risk for the
project. More the number of hazards higher is the level of risk. Once the
Machine Learning for Risk Analysis 39

FIGURE 3.2
Components of risk analysis.

identification is done, this stage is followed by risk perception, which


means acceptance of the risk identified. How the risk is perceived varies
from person to person and also upon the characteristics of the risk. If the
client wants the project to be successful and earn a profit, then he/​she
is bound to accept the risk and also take appropriate measures in order
to minimize it. In order to cure the problem first, we need to know the
problem; similarly, in order to remove the risks first, the hazard needs
to be identified as well as accepted and not ignored.
(ii) Risk assessment: When the level of risk increases, it becomes necessary
to assess these risks in order to find solutions at the right time. This is
done to ensure the safety of the project as well as the workers and users
(Paltrinieri, 2015).
(iii) Risk management: This refers to managing the risks that are being
assessed and analyzed. Different strategies are used to manage the
risks. These strategies aim to remove the negative effects of the risk and
increase productivity (Pasman, 2014).
(iv) Risk communication: This refers to making this process interactive
between the clients, developers, and users by exchanging and transfer-
ring the information related to risk with each other. This stage involves
the active participation of all those who are associated with a project
directly or indirectly, that is, from the lower to higher level in order
to make the process efficient and fulfill the demands of all stages. It is
important because it makes the decision-​making process easier as the
decision involving the lowest risk probability is chosen. It is done by
discussing and analyzing different hazards and their impact. It reduces
the gap between the expectations and results provided. Also, it gives
voice to the users to address their disappointments, which are also kind
of a risk (Aven 2014; Bucelli, 2017).
40 Machine Learning Algorithms and Applications in Engineering

3.2 Risk Assessment and Risk Management


When the level of risk increases, it becomes necessary to assess these risks in
order to find the solutions at the right time. This is done to ensure the safety
of the project as well as the workers and users. For example, mechanical pro-
ject risks can be dangerous for workers, which are technology-​related project
risks that can prove to be dangerous for users’ data. To assess the hazards,
one must figure out the origin, consequences, and the results. To identify the
origin, all the sources must be assessed to understand where things could
go wrong. Then, it has to be figured out how much are these risks exposed
to the outside environment. This stage is known as risk estimation in which
individual systems are assessed to know the contributions of each source and
functionality toward the risk (Oien, 2011; Noh, 2014; Nyvlt, 2015).
Understanding the individual systems helps in forming connections and
links of the sources with the risk being produced during development. If the
risk is assessed during the development process, it can be easily cured and
removed (Nyvlt, 2012). It also helps to predict which features can produce
failures in future and they can be continuously monitored and managed.
Hence the qualitative analysis of risk assessment includes a detailed study of
logical flow of data and the system functionalities.
Generally, the risks that are assessed are represented in the form of diagrams,
flow charts, and decision trees for better understanding. For risk assessment,
some factors such as the equipment used, the software developed, the func-
tionalities produced, and target audience must be kept in mind. This step
is very important in order to build the gap between the risk that is being
predicted and that is being perceived (Paltrinieri, 2017). This is because the
risk predicted is generally statistical in nature and is based on the theory of
probability and number of assumptions. So, in order to bridge this gap, a
quantitative and numerical approach must be adapted rather than estimating
the factors on the basis of their frequency. This stage of analyzing the predic-
tion with the result is called risk evaluation and is a part of risk assessment.
Different tests are performed and various test drives are run to assess and
predict the risk so that the final product is safe and successful. Also, various
data sets are created in order to keep track of failures produced in the past
and to ensure that same hazards are not produced in future.

3.2.1 Risk Management
The next step is to manage the risks that are being assessed and analyzed.
Different strategies are used to manage the risks. These strategies aim to
remove negative effects of the risk and increase productivity. Some people
tend to manage the risk by avoiding it, but it is not the correct way to deal
with the hazards. Because if they are not dealt with at the earlier stage, they
will grow into something serious. So, the correct way of managing risks is
Machine Learning for Risk Analysis 41

first of all accepting them (Paltrinieri, 2019). Risk management is done on the
basis of priority, which means the risk with highest probability of occurrence
or the risk that can cause maximum amount of damage needs to be dealt with
first and the risks that are expected to cause lower loss are prioritized in the
descending order. Hence, a classification of the risks needs to be done on the
basis of loss and probability.
Generally, the amount of loss faced is given more priority than the fre-
quency of its occurrence. It is the most important step because if the risks
are not managed properly, then there is no point in identifying and assessing
them. The unmanaged risks reduce the rate of profit, durability, quality, and
reliability of the project. It can also affect the brand value of a company and
decrease the trust of users. The problem or limitation faced by this phase
is the limited amount of technique and resources available to manage the
hazards and the increase in cost due to risk management. The steps required
in risk management are:

(i)
Identification of the risk and its resource and domain
(ii)
Considering the impact of this on the system
(iii)
Probability of occurrence
(iv) Impact on effective cost
(v) Consequences on the project
(vi) Classification on the basis of priority
(vii) Assessing the constraints that can be faced
(viii) Describing the needs of users and agenda for doing this activity
(ix) Engaging in discussions and communication to decide the managing
technique
(x) Finally, developing an analytical approach in order to manage the risk
(xi) Organizing the resources and cost required for the managing process

Thus, risk management is done on the basis of results obtained from risk
assessment and then the appropriate techniques are chosen by managers.

3.3 Risks in the Business World


As we all know, any business, small or big, is prone to risk. Some of them
even cause loss of profits to the extent of bankruptcy. The large firms have
expanded their sections of risk management. The smaller firms or businesses
are still not able to find systematic ways to handle the issue. Comprehensive
business plans are required to have a successful business, but the plans need
to be updated with time (Svozil, 1997). With changes in the market and its
42 Machine Learning Algorithms and Applications in Engineering

standards, the old methods of risk management and analysis require modifi-
cation. The major risks that the firms face are the following:

(i) Strategic risk: As the name suggests, strategic risk is strategy-​related, in


which the planning is not appropriate for the successful development
of the business. This is due to the technological changes in the society,
new competitors with modern technologies entering the market, rapid
changes in the customer demand section, and an increase in the cost of
raw material. Studies show the example of firms that suffer strategic
risk. Some trained their models to gain success, and some did not. An
example of the same is Kodak, which had a decent position in the pho-
tography market. When engineers developed the digital camera, it was
as a threat to its central marketing model, and Kodak failed to modify
it. If Kodak had considered the strategic risk more thoroughly, it could
have led to a better future for the company. Handling a strategic risk is
not always destructive as there are examples of brands that were able
to see it as a chance of change like Xerox. It became a success with a
single and satisfactory product, that is, a photocopy machine. The com-
pany tried to add a laser printer to enhance the business model and was
ready to face the statistical risk. The company saw the positive side of
the change due to good planning.
(ii) Compliance risk: Any company always tries to increase their profits and
business prospects, but they have to comply with law while doing the
same. The laws constantly change with time and regulation is required.
For example, a person makes cosmetic products in his factory and sells it
across the country. The person is doing so well in the industry and thus
wants to expand the business to another country. Now the business is
under the compliance risk. The countries have some rules regarding the
import of products and these rules and regulations may cost money and
may not give a bigger margin in profits in comparison to the expected
margins. The same condition may arise if the product line is changed.
(iii) Operational risk: This type of risk is due to the unexpected failure in the
way the company operates. The risks include mechanical failure, tech-
nical failure, or failure due to people. If a person is to be given a salary of
₹10,000 and is handed a check for ₹1,000, then it is a failure of process and
people. This may look like a small problem in the example but in the big
industry, it can cause issues with bigger amounts. Other problems like
natural disasters and power cuts can cause operational damage. From
these examples, operational risk can be defined as any issue or risk that
damages the function or operation of a business (Villa, 2016a, 2016b).
(iv) Financial risk: Most of the times, the risk that the firm faces is categorized
as financial risk, that is, a risk having a financial issue leading to low rev-
enue and decreased profits. Let us assume that a client has the credit exten-
sion of 30 days and is a large source of the company’s revenue (Marchi,
1999). The company, in this case, is under financial risk as a great deal of
Machine Learning for Risk Analysis 43

trouble can be caused if the payment is not done on time. Similar is the
case when a particular company is in great debt. International businesses
undergo huge financial risks. We can go back to the example of cosmetics
where the person wanted to sell the products in the international market
and if the person tries to sell the goods in the United States, UK, France,
and India, the company may have to bear the conversion charges of these
different currencies. This type of example comes under financial risk.
(v) Reputational risk: One thing that is common to all the businesses is their
reputation, be it a small size business or a multinational firm. If the
reputation is positive, then the selling of commodity and recruitment of
employees become easier. If the reputation is damaged, then there is an
immediate loss in revenue and the employees might leave the company
too. The advertising agencies may not show interest in the work related
to the company and so on (Yang, 2015).

3.3.1 Models in the Field of Risk Management


Various models have been proposed for managing risks in different scenarios.
Some of the models are explained as follows:

(i) Interpretability: Machine learning designs have a character of signifying


results that is hard to understand or evaluate. To increase the interpret-
ability, some models are used as follows:
(a) Linear monotonic models: Linear coefficients act as a huge support for
the exhibition of the output. Linear regression models come under
this category.
(b) Nonlinear monotonic models: Restraining data so that we can see
either a falling or a rising global relationship where the variables
are simplified for producing the result. Gradient boosting model is
an example of the same.
(c) Nonlinear no monotonic model: Methodologies such as local interpret-
able model-​agnostic explanations or Shapley values help ensure
local interpretability. The examples include models like uncon-
strained deep learning models.
(ii) Bias: This feature takes care of the trueness of the system. The valid-
ation takes care of the proper implementation of rules for fairness in
the mechanism. Four approaches are generally used depending on the
requirements:
(a) Demographic blindness: Settlements are composed utilizing a
restricted assortment of characteristics that are related to the
preserved classes.
(b) Demographic parity: For each protected class, results are proportion-
ally the same.
44 Machine Learning Algorithms and Applications in Engineering

(c) Equal opportunity: For each protected class, true positive standards
are similar.
(d) Equal odds: For each protected class, true-​positive and false-​positive
rates are equal.
(iii) Feature engineering: It is the method of managing data to create
innovations that can be executed with the help of machine learning
algorithms. It is the process of selecting significant characteristics from
a fresh pool of data and converting them into forms that are fit for
machine learning. Feature engineering model development in ML is
complicated in comparison to the models used traditionally. The first
reason for this is that machine learning models can combine a higher
amount of information. The second is that there is a need for feature
engineering that is required for disorganized data references at the pre-​
processing level before the training method can start.
(iv) Hyper parameters: The variables which define the system composition
and decide the network training method are the hyper parameters. It is
very important to understand the variables and determine the appro-
priate selection of hyper parameters. The approaches for the selection of
hyper parameters include the latest practices used in the industry and
expert judgment.
(v) Production readiness: Machine learning models, despite being algo-
rithmic, require a lot of computation. This element is generally viewed
in the process of model development. The validation is done previously
to evaluate a variety of model risks connected with its usage and for
machine learning, they expand their scope. There is a requirement of
setting a limitation on the data flow through the model, keeping in
mind the run time and the architecture of the model.
(vi) Dynamic model calibration: Sometimes, there is a dynamic change in the
parameters in some types of models depicting the data patterns. The
validators can easily decide which dynamic calibration is best suited
for the firm. The factors that are evaluated include the development of
a monitoring plan, ensuring proper control to reduce risk in accordance
with the usage of the model.

3.4 Machine Learning Techniques for Risk Management


and Assessment
Two types of ML techniques are used in risk analysis. In a supervised
ML technique, the input is pre-​decided under supervised conditions and
generates an expected output (Wolpert, 2002; Jain 2020). In an unsupervised
Machine Learning for Risk Analysis 45

ML technique, various kinds of inputs are given in undefined sets and an


unexpected output is generated.

(i) Supervised technique: It is further divided into regression and classifica-


tion. Regression technique is used to define the relationship between
the variable assigned to the risks predicted during the assessment.
For example, a regression equation for a credit-​based assessment can
be defined as the variable assigned to non-​payment of the loan taken,
which is further analyzed and sorted using a long range of variables
that are independent and are assigned to functions such as employment
condition of the person, past payment history, and property.
This method is very useful as the independent variables facilitate
in the management process and it is much better than the traditional
method as it is self-​explanatory and less theoretical and more prac-
tical. This method is suitable for big data sets as the variables that
are not in use are deleted automatically. The other main technique is
the classification technique. It is the most commonly used technique
in which risks are prioritized and represented in the form of decision
trees and data flow diagrams. This method is easy and more compre-
hensive due to the visual representation and hence requires less time
in management.
This model is understood by both technical and nontechnical
members. The risks are branched into various categories according to
the percentages, which indicate the probability of the occurrence of the
various risks and hazards. This categorization and subdivision tech-
nique helps to determine the later consequences as well. This makes the
sorting possible in less time as it contains only decision based yes/​no
type questions. The only limitation of this module is that its predictive
power is low, which means it helps in analyzing but not in predicting
when compared to the traditional methods.
(ii) Unsupervised technique: This method is further categorized into clustering.
This is also based on subdivision, but it forms similar subgroups and
does not predict any output. This is a kind of a grouping technique with
no prediction function. The example of this technique is spam detection
in emails. This means all the similar looking mails form a cluster that is
if a mail is considered spam, then all the similar looking mails will be
clustered as spam. This assignment of variables is lacking, which does
not allow it to detect or predict anything. This just helps the analysts to
figure out the patterns and similarities between the functions.

The ML techniques form various models that are quite complex in nature as
it uses too many variables and other parameters in comparison of the output
generated, which make it heavy as well as complex as shown in Figure 3.3.
Due to the complexity in design, the interpretation time also increases, which
46 Machine Learning Algorithms and Applications in Engineering

FIGURE 3.3
Machine learning assessment model techniques.

is undesirable. A good predictive model is generally hard to analyze. For effi-


cient risk assessment and management, interpretation is as important as pre-
diction. So, some methods can be applied on these models to overweight
the observations in comparison to parameters or inputs so that more and
more outputs can be generated. Such a method is called “boosting.” Another
way is running a model on different data sets many times so that it produces
different predictions for the same dataset. This is called bagging. The final
desirable model is the average of boosting and bagging so that it can both
predict as well as interpret at the same time.
In order to remove all the complexities, machine learning is combined with
some techniques of deep learning, which lead to advancement in technology.
These are generally referred to as the practical approach or theory-​ free
approach as it does not involve much of an explanation as they are mostly
self-​explanatory. In deep learning, all the algorithms are equipped in such a
way that they represent a particular aspect of data and enhance the predic-
tion of that part of data. This is how the data can be easily predicted by using
Ml techniques and interpreted with a little help of deep learning techniques
without any increase in the complexity. This allows easy interpretation of
even low quality and unstructured data, which is present in huge data sets
and makes the understanding easy for both analysts and nontechnical users.
This integrated approach allows all the features to run together as system and
perform a relevant analysis. Various ML algorithms applied by distinguished
researchers to solve various risks is explained in Table 3.1.

3.4.1 Challenges of Machine Learning in Risk Management


Apart from all the use cases and advantages of using ML in risk analysis and
management, there exist some of the major issues that need to be taken into
Machine Learning for Risk Analysis 47

TABLE 3.1
ML Algorithms for Risk Management

Risk Type ML Algorithms

Credit risk Neural Networks, SVM, KNN, Random Forest, Lasso regression,
Cluster analysis
Liquidity risk SVM, ANN, Bayesian Networks
Market risk GELM, Cluster analysis, SOM, Gaussian Mixtures, cluster analysis
Operational risk Nonlinear clustering method, Neural Networks, k-​Nearest Neighbor,
Naïve Bayesian, Decision Tree.
RegTech risk SVM

consideration. Some of the challenges of machine learning in risk manage-


ment that the industry faces are as follows:

(i) Availability of data: One of the most significant challenges is to obtain


relevant as well as suitable data for processing. There is a lot of poten-
tial for machine learning packages for performing multiple tasks such
as processing images and natural language by reading all types of data,
but it is becoming difficult to maintain discipline in the data internally.
Moreover, the sharing of data is tough as the information is stored in
different systems with different restrictions that make the information
retrieval even more challenging. Also, perhaps the information may not
be stored formally.
(ii) Availability of skilled staff: With the evolution of new techniques, the
requirement of skilled labor is quite demanding. The implementation,
understanding, and working of advanced features and solutions is of
high importance. Since providing training to the unskilled staff is time-​
consuming, there are some attempts made to solve this problem by
offering a course to the aspirants by building a campus with around
7,000 people in India where there is frequent presence of these skills.
(iii) Accuracy of machine learning techniques: It is not as simple as just
applying various machine learning techniques to manage the potential
risk; rather, it requires continuous evaluation of the process in order to
obtain optimal current solution. In order to ensure feasibility, multiple
testing of the technique is required.
(iv) Transparency and ethics: This is one of the popular issues that usually
occur. It is evident from the fact that the models work between the
layers and remain hidden in the initial input and final output. This
can lead to compliance problems while validating the model. It may
even turn into the loss of a firm or convergence of the uniform ideal
for trading, resulting in a lot of risk. There are many ethical issues as
well such as unequal lending decision due to gender, race, and sexual
discrimination.
48 Machine Learning Algorithms and Applications in Engineering

3.4.2 Machine Learning Use Cases in the Financial Sector


This section analyzes and gives a detailed description of various use cases of
machine learning in the field of risk management in the financial sector such
as fraud detection, credit risk, supervision of conduct, and market abuse in
trading (Tanwar, 2020).

(i) Fraud detection: Machine learning is now often used in the field of fraud
detection, mainly to detect credit card fraud in the banking system.
It has resulted in higher significant success rate till now. Banks have
set up various monitoring and surveillance systems in order to ensure
security. These systems keep a track on the payment fraud activities
that take place more often. The fraud model engine works on the trad-
itional historical payment algorithm that blocks the fraudulent trade as
soon as it detects it.
The training, testing, and authorization of machine learning
algorithms become possible because of large data sets provided by
the credit card payments. Moreover, various classification algorithms
are trained with the help of historical data with identified fraud and
nonfraud tags. This training is done with the help of large historical
nonfraudulent data. The above historical payment data sets generate a
clear view of features of the card by distinguishing them on the basis of
transactions, owner, and history of the payment.
(ii) Credit risk: The term “credit risk” signifies the prospects of loss because
of unsuccessful payments made by the lender. Therefore, the credit risk
management system (CRMS) is implemented in order to meet the losses
by analyzing the fairness of the bank and its capital along with the loan
loss section at any instant of time. This system has enhanced the trans-
parency as demanded by the financial crises. Transparency is carried by
paying more emphasis on the regular examination of the knowledge of
banks for its customers and credit risks.
Suitable credit risk management improves the entire performance
of the system and provides a competitive benefit. Various prediction
models are made in order to make predictions regarding the kind of
lender. This is where the machine learning techniques are implemented.
These machine learning algorithms provide positive results and greater
success rates in solving the problems including credit risks as well.
(iii) Supervision of conduct and market abuse in trading: Another use case of
machine learning in the risk management system is the surveillance of
the performance loopholes generated mainly by the traders working in
the financial institutions. The various trading illegalities often lead to
economic as well as status failure. In order to overcome such crimes and
shortcomings, numerous self-​operating systems have been developed
in order to provide a check on the trading behaviors of dealers with
increased accuracy and distinct ways to identify them.
Machine Learning for Risk Analysis 49

Earlier, in the initial stages, the system’s performance was limited


to checking the behavior of a single trade. But with time, various new
machine learning approaches are applied and the machines can now
identify huge and complicated data and analyze the entire portfolios.
Apart from performing this analysis, the systems are also able to link
distinct trading facts via email flux, calendar items, developing check-​
in and check-​out times, and calls as well. The automatic analysis is
performed using deep learning and text mining techniques, which
make the system machine readable.

3.5 Case Studies to Understand the Role of Machine


Learning in Risk Management

i)  Case Study 1: Banks

On average banks are prone to various risks such as credit risk, market
risk, foreign exchange risk, foreign risk, frauds, etc. These risks are treach-
erous for the reputation of the banks and their worthwhile procedures.
These risks are dormant privation to the banks and they can be
questioned for the obligations on them. Risks are unpredictable and can
develop irrespective of time. How well would it be if the risk can be spe-
cified or predicted? To specify the upcoming risks in advance, machine
learning can be implemented. All the possible risks are first analyzed and
reviewed. During the process of analysis of these risks a large amount of
data is collected which is basically unstructured in nature. This data is
collected from market information, customer reviews, metadata etc.
Machine learning is all about working with data. It processes this data
for working machines. The machines will produce a desired output on
the basis of the input processed. When machine learning along with some
algorithms will be implemented on the large amount of unstructured
data. The machine will interact on the basis of the data and the output
will be generated. This generated output can be a sort of prediction of the
upcoming risk that might rise up in the near future.
For the accurate results from the system, it is important the training of
the large amount of unstructured data that is collected is done in a perfect
way. After the training is finished the data is converted into structured
data or the labeled data. On this final set of data various algorithms
are processed. These algorithms basically mould the data into a frame-
work of a process to take place on the basis of which the final output is
generated. The type of data collected in the final output will also be of the
same category.
50 Machine Learning Algorithms and Applications in Engineering

FIGURE 3.4
Role of machine learning in risk management for disaster management.

Machine learning is compatible, dynamic and can prevent a bank from


such cunning risks. A number of problems can be easily fathom with the
help of machine learning in the fields of risk assessment.

ii)  Case Study 2: Disaster management

Disaster can be defined as the natural affliction that can lead to various
human, environmental and property losses. These disasters are unpre-
dictable but there are several technologies which are operated today to
predict an approximate time for such calamities. The labeled data which
is collected for machine learning is used to predict the disaster in the form
of a sequence of their occurrence. There are various processes which can
forecast the occurrence of the disasters on the basis of their probability
of occurrence as shown in Figure 3.4. During these processes there is a
chance that several fallacies occur. This measure indicates misfiring of
appliance or apparatus, preservation excess, quantity of supplementary
worked, etc.
These machines that predict the upcoming disasters and the human
generation cannot afford risks in these indicators as everyone relies on
their final output. Consequently, the role of risk assessment in fields
like disaster management becomes necessary for protection of a specific
area from sudden tragedy and irreplaceable losses. Machine learning
includes several techniques that look after the prevention of such risks
in this area.

Apart from all the issues and problems, looking at the positives in a situation
is a much more crucial segment. Considering the future of machine learning
in risk management is worthy. This is evident from the fact that the cost of
development is reduced. Moreover, the time-​consuming nature of this pro-
cess has also decreased at a higher scale. One such instance is that of Banco
Bilbao Viscera Argentaria (BBVA), a financial service company in Spain. This
Machine Learning for Risk Analysis 51

multinational corporation is investing heavily in compliance-​based workers.


Therefore, various machine learning techniques are used in order to reduce
this outlay. This technology makes the repetitive tasks automatic. Along with
the organization, data clustering as well as recovering of advanced data are
some of the abilities of machine learning that will provide a huge benefit to
the companies.

3.6 Conclusion
Throughout the chapter, various aspects and techniques of machine
learning that are engaged in the fields of risk assessment to ameliorate this
process has been discussed. The importance of machine learning in risk
assessment is skillfully acknowledged. Machine learning has the poten-
tial to create wonders in risk assessment by incorporating methodical
procedures and models that can fabricate accurate results for risk analysis
by effectively monitoring large and complex datasets. The application of
machine learning in this sector has led to the conclusion that these methods
or techniques can be used to analyze huge amounts of data with efficient
predictive analysis. Various use cases have also been discussed such as
supervision of conduct and market abuse in trading, fraud detection, and
credit risk.
The description of various models and the case studies that are covered in
the chapter is a blueprint of the utilization of machine learning in different
industries and organizations in the fields like risk management in banks
and risk management in disaster management. The major issue that can be
addressed by employing technology is to counter risks as machine learning
provides various models and techniques that can minimize or prevent such
risks. These techniques and models change with different sets of labeled
data as per the convenience of the procedure. It provides the stability in
the field and aids in the protection of human and national information
that can be considered as highly confidential information. Still research is
ongoing in the area of risk assessment and several measures are already
implemented. There are still many areas untouched by machine learning. It
can ease the stress of several industries and organizations as potential risks
can be identified and suitable methods can be employed if the probability
of risk is strong.
In the end, one can just say that humans will be encountering an era which
will make even complex problems easy to solve with efficiency and they will
solve it for sure, which will be time-​saving and will be economically benefi-
cial. Time is not far afar when technology such as AI and ML is employed as
a solution to almost every problem.
52 Machine Learning Algorithms and Applications in Engineering

References
Apostolakis, G.E. 2004. How useful is quantitative risk assessment? Risk Analysis
24: 515–​520.
Aggarwal, P.K., Grover, P.S., Ahuja, L. 2018. Exploring quality aspects of smart mobile
phone applications. Journal of Advanced Research in Dynamical and Control Systems
11: 292–​297.
Aggarwal, P.K., Grover, P.S., Ahuja, L. 2019. Assessing quality of mobile applications
based on a hybrid MCDM approach. International Journal of Open Source Software
and Processes, 10: 51–​65.
Aven, T. 2012. The risk concept—​historical and recent development trends. Reliab.
Eng. Syst. Saf. 99: 33–​44.
Aven, T., Krohn, B.S. 2014. A new perspective on how to understand, assess and
manage risk and the unforeseen. Reliab. Eng. Syst. Saf. 121:1–​10.
Bucelli, M., Paltrinieri, N., Landucci, G., 2018. Integrated risk assessment for oil and
gas installations in sensitive areas. Ocean Eng. 150: 377–​390.
Chen, H., Moan, T., Verhoeven, H. 2008. Safety of dynamic positioning operations on
mobile offshore drilling units. Reliab. Eng. Syst. Saf. 93: 1072–​1090.
Cheng, H.-​T., Koc, L., Harmsen, J., Shaked, T., Chandra, T., Aradhye, H., Anderson,
G., Corrado, G., Chai, W., Ispir, M. 2016. Wide and deep learning for recom-
mender systems. In: Proceedings of the 1st Workshop on Deep Learning for
Recommender Systems. ACM, pp. 7–​10.
Comfort, L.K. 2019. The Dynamics of Risk: Changing Technologies, Complex Systems, and
Collective Action in Seismic Policy. Princeton, NJ: Princeton University Press.
Creedy, G.D. 2011. Quantitative risk assessment: how realistic are those frequency
assumptions? J. Loss Prev. Process Ind. 24: 203–​207.
De Marchi, B., Ravetz, J.R. 1999. Risk management and governance: a post-​normal
science approach. Futures 31: 743–​757.
Durga Rao, K., Gopika, V., Sanyasi Rao, V.V.S., Kushwaha, H.S., Verma, A.K., Srividya,
A. 2009. Dynamic fault tree analysis using Monte Carlo simulation in probabil-
istic safety assessment. Reliab. Eng. Syst. Saf. 94: 872–​883.
Diekmann, E.J. 1992. Risk analysis: lessons from artificial intelligence. Int. J. Proj.
Manag. 10, 75–​80.
Goodfellow, I.J., Bengio, Y., Courville, A. 2016. Deep Learning. The MIT Press,
Cambridge, MA.
Hastie, T., Tibshirani, R., Friedman, J. 2009. Unsupervised learning. In: The Elements of
Statistical Learning. Springer, Berlin, pp. 485–​585.
Haugen, S., Vinnem, J.E. 2015. Perspectives on risk and the unforeseen. Reliab. Eng.
Syst. Saf. 137: 1–​5.
Jain, P., Sharma, A., Ahuja, L. 2018. The impact of agile software development pro-
cess on the quality of software product. In: Proceedings of the International
Conference on Reliability, Infocom Technologies and Optimization (Trends and
Future Directions) (ICRITO), pp. 812–​815.
Jain, P., Sharma, S. 2019. Prioritizing factors used in designing of test cases: An ISM-​
MICMAC based analysis. In: Proceedings of International Conference on Issues
and Challenges in Intelligent Computing Techniques (ICICT), India.
Machine Learning for Risk Analysis 53

Jain, P., Sharma, A., Aggarwal, P.K. 2020. Key attributes for a quality mobile applica-
tion. In Proceedings of the International Conference on Cloud Computing, Data
Science and Engineering Confluence, 50–​54.
Kaplan, S., Garrick, B.J., 1981. On the quantitative definition of risk. Risk Anal.
1: 11–​27.
Khakzad, N. 2015. Application of dynamic Bayesian network to risk analysis of
domino effects in chemical infrastructures. Reliab. Eng. Syst. Saf. 138: 263–​272.
Khakzad, N., Khan, F., Amyotte, P. 2013a. Risk-​based design of process systems using
discrete-​time Bayesian networks. Reliab. Eng. Syst. Saf. 109: 5–​17.
Khakzad, N., Khan, F., Amyotte, P. 2013b. Quantitative risk analysis of offshore
drilling operations: a Bayesian approach. Saf. Sci. 57: 108–​117.
King, G., Zeng, L. 2001. Logistic regression in rare events data. Polit. Anal. 9: 137–​163.
Kongsvik, T., Almklov, P., Haavik, T., Haugen, S., Vinnem, J.E., Schiefloe, P.M. 2015.
Decisions and decision support for major accident prevention in the process
industries. J. Loss Prev. Process Ind. 35: 85–​94.
Landucci, G., Paltrinieri, N. 2016a. A methodology for frequency tailorization
dedicated to the Oil & Gas sector. Process Saf. Environ. Prot. 104: 123–​141.
Landucci, G., Paltrinieri, N. 2016b. Dynamic evaluation of risk: from safety indicators
to proactive techniques. Chem. Eng. Trans. 53: 169–​174.
Lasi, H., Fettke, P., Kemper, H.-​G., Feld, T., Hoffmann, M. 2014. Industry 4.0. Bus. Inf.
Syst. Eng. 6: 239–​242.
Musgrave, G.L. 2013. James Owen Weatherall, the physics of Wall Street: A brief his-
tory of predicting the unpredictable. Bus. Econ. 48: 203–​204.
Nivolianitou, Z.S., Leopoulos, V.N., Konstantinidou, M. 2004. Comparison of
techniques for accident scenario analysis in hazardous systems. J. Loss Prev.
Process Ind. 17: 467–​475.
Nobre, F.S., 2009. Designing Future Information Management Systems. IGI Global.
Noh, Y., Chang, K., Seo, Y., Chang, D. 2014. Risk-​based determination of design
pressure of LNG fuel storage tanks based on dynamic process simulation
combined with Monte Carlo method. Reliab. Eng. Syst. Saf. 129: 76–​82.
Nývlt, O., Haugen, S., Ferkl, L. 2015. Complex accident scenarios modelled and
analysed by stochastic Petri nets. Reliab. Eng. Syst. Saf. 142: 539–​555.
Nývlt, O., Rausand, M. 2012. Dependencies in event trees analyzed by Petri nets.
Reliab. Eng. Syst. Saf. 104: 45–​57.
Øien, K., Utne, I.B., Herrera, I.A. 2011. Building safety indicators: Part 1 –​theoretical
foundation. Safety Sci. 49: 148–​161.
Paltrinieri, N., Khan, F., Cozzani, V. 2015. Coupling of advanced techniques for
dynamic risk management. J. Risk Res. 18: 910–​930.
Paltrinieri, N., Reniers, G. 2017. Dynamic risk analysis for Seveso sites. J. Loss Prev
Process Ind. 49: Part A: 111–​119.
Paltrinieria, N., Comfortb, L., Reniers, G. 2019. Learning about risk: Machine learning
for risk assessment. Safety Sci. 118: 475–​486.
Pasman, H., Reniers, G. 2014. Past, present and future of quantitative risk assessment
(QRA) and the incentive it obtained from land-​use planning (LUP). J. Loss Prev.
Process Ind. 28: 2–​9.
Svozil, D., Kvasnicka, V., Pospichal, J. 1997. Introduction to multi-​layer feed-​forward
neural networks. Chemom. Intell. Lab. Syst. 39: 43–​62.
54 Machine Learning Algorithms and Applications in Engineering

Tanwar, S., Bhatia, Q., Patel, P., Kumari, A., Singh, P.K., Hong, W.C. 2020. Machine
learning adoption in blockchain-​based smart applications: The challenges, and
a way forward. IEEE Access. 8: 474–​488.
Villa, V., Paltrinieri, N., Khan, F., Cozzani, V. 2016a. Towards dynamic risk analysis: a
review of the risk assessment approach and its limitations in the chemical pro-
cess industry. Saf. Sci. 89. https://​doi.org/​10.1016/​j.ssci.2016.06.002
Villa, V., Paltrinieri, N., Khan, F., Cozzani, V., 2016b. A short overview of risk analysis
background and recent developments. In: Dynamic Risk Analysis in the Chemical
and Petroleum Industry: Evolution and Interaction with Parallel Disciplines in
the Perspective of Industrial Application. 10.1016/​B978-​0-​12-​803765-​2.00001-​9.
Wolpert, D.H., 2002. The supervised learning no-​free-​lunch theorems. Soft Comput.
Ind. 25–​42.
Yang, X., Haugen, S., 2015. Classification of risk to support decision-​making in haz-
ardous processes. Safety Sci. 80: 115–​126.
newgenprepdf

4
Machine Learning Techniques
Enabled Electric Vehicle

Shyamalagowri Murugesan1 and Revathy Jayabaskar2


1
Associate Professor, Department of Computer Science Engineering,
Erode Sengunthar Engineering College (Autonomous), Erode, India
2
School of Computing, SASTRA Deemed University, Thanjavur, India
E-​mail: [email protected]; [email protected]

CONTENTS
4.1 Introduction................................................................................................... 56
4.1.1 Artificial Intelligence Technology to Enhance EV
Production and Support the Deployment of
Electric Vehicles................................................................................ 56
4.1.2 Artificial Intelligence Used to Supercharge Battery
Development for Electric Vehicles................................................. 58
4.1.3 A Smarter Approach to Battery Testing........................................ 58
4.1.4 Wider Applications.......................................................................... 59
4.1.5 A Review on AI-based Predictive Battery Management
System for E-Mobility...................................................................... 60
4.2 Reverse Engineering with AI for Electric Power Steering...................... 61
4.3 Artificial Intelligence Technology to Enhance EV Production and
Support the Deployment of Electric Vehicles........................................... 62
4.4 Artificial Intelligence Used to Supercharge Battery
Development for Electric Vehicles............................................................. 63
4.5 How AI Helps Build Better Batteries......................................................... 65
4.6 Machine Learning Supercharges Battery Development......................... 67
4.7 AI-Based Predictive Battery Management System for E-Mobility........ 67
4.8 Uses of Artificial Intelligence for Electric Vehicle
Control Applications.................................................................................... 69
4.9 Conclusion..................................................................................................... 71

DOI: 10.1201/9781003104858-4 55
56 Machine Learning Algorithms and Applications in Engineering

4.1 Introduction
Machine learning is expected to play a vital role in the upcoming industry
revolution. The evolution of ML and AI has great implications for the devel-
opment of electric vehicles in various means. Battery performance can make
or break the electric vehicle experience, from driving range to charging time
to the lifetime of the car. Now, artificial intelligence has made dreams like
recharging an EV in the time it takes to stop at a gas station a more likely
reality and could help improve other aspects of battery technology.
For decades, advances in electric vehicle batteries have been limited by a
major bottleneck: evaluation times. At every stage of the battery develop-
ment process, new technologies must be tested for months or even years to
determine how long they will last. But now, a team led by Stanford professors
Stefano Ermon and William Chueh has developed a machine learning-​based
method that slashes these testing times by 98 percent. Although the group
tested their method on battery charge speed, they said it can be applied
to numerous other parts of the battery development pipeline and even to
nonenergy technologies.
“In battery testing, you have to try a massive number of things, because
the performance you get will vary drastically,” said Ermon, an assistant pro-
fessor of computer science. “With AI, we’re able to quickly identify the most
promising approaches and cut out a lot of unnecessary experiments.”

4.1.1 Artificial Intelligence Technology to Enhance EV Production


and Support the Deployment of Electric Vehicles
Electric vehicles reduce tailpipe emissions from the transportation sector and
produce important public health benefits. One of the most important comple-
mentary assets for EV adoption is charging station infrastructure.
While we initially thought that the availability of real-​time transactions
from charging stations would be a primary innovation for this research, we
quickly realized that we also had an enormous quantity of unstructured
text to learn about how EV users engage with each other. Surprisingly
until this point, the best evidence on EV consumer behavior typically
relied on surveys or simulated data, and these were limited to a single
market or region. We understood that data from mobile platforms offered
a unique ability to aggregate large-​scale evidence in ways not previously
possible.
What was most exciting for our research team was the fact that the data
was streaming and offered the potential to update in near real-​time. This is in
stark contrast to data collection from large government surveys, which take
several years to complete, and data discovery is often slow and expensive.
So, we set off to see if we could analyze both government-​run and privately
run EV charging stations to provide national, evidence-​based analysis.
Machine Learning Techniques Enabled Electric Vehicle 57

Given the quantity of data however, a major challenge for the team was the
fact that it would take human experts about 32 weeks to sift through EV user
reviews in order to extract useful insights from free text.
So, almost on the side, we started experimenting with deep learning nat-
ural language processing techniques to unlock some insights there. It turned
out that the review data was an untapped source of research innovation for
us. We quickly realized that the application of AI/​ML to this data could both
accelerate policy analysis and reduce science and technology (S&T) research
evaluation costs.
Consequently, by deploying deep learning techniques to analyze those
EV user reviews, we were able to show how machine learning tools could
be used to quickly analyze streaming data for policy evaluation in near
real time and provide new insight into important electric vehicle charging
infrastructure policies, such as the need to focus on the quality of the user
experience and evidence supporting public involvement in EV charging
network buildout.
By displacing gasoline and diesel fuels, electric cars and fleets reduce
emissions from the transportation sector, thus offering important public
health benefits. However, public confidence in the reliability of charging
infrastructure remains a fundamental barrier to adoption. Using large-​scale
social data and machine learning from 12,720 electric vehicle (EV) charging
stations, we provide national evidence on how well the existing charging
infrastructure is serving the needs of the rapidly expanding population of
EV drivers in 651 core-​based statistical areas in the United States. We deploy
supervised machine learning algorithms to automatically classify unstruc-
tured text reviews generated by EV users. Extracting behavioral insights at
a population scale has been challenging given that streaming data can be
costly to classify. Using computational approaches, we reduce processing
times for research evaluation from weeks of human processing to just
minutes of computation. Contrary to theoretical predictions, we find that
stations at private charging locations do not outperform public charging
locations provided by the government. Overall, nearly half of the drivers
who use mobility applications have faced negative experiences at EV char-
ging stations in the early growth years of public charging infrastructure, a
problem that needs to be fixed as the market for electrified and sustainable
transportation expands.
The study, published by Nature on February 19, 2020, was part of a larger
collaboration among scientists from Stanford, Massachusetts Institute of
Technology (MIT), and the Toyota Research Institute that bridges foun-
dational academic research and real-​ world industry applications. The
goal: finding the best method for charging an EV battery in 10 minutes that
maximizes the battery’s overall lifetime. The researchers wrote a program
that, based on only a few charging cycles, predicted how batteries would
respond to different charging approaches. The software also decided in real
time what charging approaches to focus on or ignore. By reducing both the
58 Machine Learning Algorithms and Applications in Engineering

length and number of trials, the researchers cut the testing process from
almost two years to 16 days.
“We figured out how to greatly accelerate the testing process for extreme
fast charging,” said Peter Attia, who co-​led the study while he was a graduate
student. “What’s really exciting, though, is the method. We can apply this
approach to many other problems that, right now, are holding back battery
development for months or years.”

4.1.2 Artificial Intelligence Used to Supercharge Battery


Development for Electric Vehicles
Battery performance can make or break the electric vehicle experience, from
driving range to charging time to the lifetime of the car. Now, artificial intel-
ligence has made dreams like recharging an EV in the time it takes to stop at
a gas station a more likely reality and could help improve other aspects of
battery technology.
For decades, advances in electric vehicle batteries have been limited by a
major bottleneck: evaluation times. At every stage of the battery develop-
ment process, new technologies must be tested for months or even years to
determine how long they will last. But now, a team led by Stanford professors
Stefano Ermon and William Chueh has developed a machine learning-​based
method that slashes these testing times by 98 percent. Although the group
tested their method on battery charge speed, they said it can be applied
to numerous other parts of the battery development pipeline and even to
nonenergy technologies.
“In battery testing, you have to try a massive number of things, because
the performance you get will vary drastically,” said Ermon, an assistant pro-
fessor of computer science. “With AI, we’re able to quickly identify the most
promising approaches and cut out a lot of unnecessary experiments.”

4.1.3 A Smarter Approach to Battery Testing


Designing ultrafast-​charging batteries is a major challenge, mainly because it
is difficult to make them last. The intensity of the faster charge puts greater
strain on the battery, which often causes it to fail early. To prevent this damage
to the battery pack, a component that accounts for a large chunk of an electric
car’s total cost, battery engineers must test an exhaustive series of charging
methods to find the ones that work best.
The new research sought to optimize this process. At the outset, the team
saw that fast-​ charging optimization amounted for many trial-​ and-​
error
tests—​something that is inefficient for humans, but the perfect problem for
a machine.
“Machine learning is trial-​and-​error, but in a smarter way,” said Aditya
Grover, a graduate student in computer science who co-​ led the study.
“Computers are far better than us at figuring out when to explore—​try new
Machine Learning Techniques Enabled Electric Vehicle 59

and different approaches—​and when to exploit, or zero in, on the most prom-
ising ones.”
The team used this power to their advantage in two key ways. First, they
used it to reduce the time per cycling experiment. In a previous study, the
researchers found that instead of charging and recharging every battery until
it failed—​the usual way of testing a battery’s lifetime—​they could predict
how long a battery would last after only its first 100 charging cycles. This is
because the machine learning system, after being trained on a few batteries
cycled to failure, could find patterns in the early data that presaged how long
a battery would last.
Second, machine learning reduced the number of methods they had to
test. Instead of testing every possible charging method equally, or relying on
intuition, the computer learned from its experiences to quickly find the best
protocols to test.
By testing fewer methods for fewer cycles, the study’s authors quickly
found an optimal ultra-​fast-​charging protocol for their battery. In addition
to dramatically speeding up the testing process, the computer’s solution was
also better—​and much more unusual—​than what a battery scientist would
likely have devised, said Ermon.
“It gave us this surprisingly simple charging protocol—​something we
didn’t expect,” Ermon said. Instead of charging at the highest current at the
beginning of the charge, the algorithm’s solution uses the highest current
in the middle of the charge. “That’s the difference between a human and a
machine: The machine is not biased by human intuition, which is powerful
but sometimes misleading.”

4.1.4 Wider Applications
The researchers said their approach could accelerate nearly every piece of
the battery development pipeline: from designing the chemistry of a battery
to determining its size and shape, to finding better systems for manufac-
turing and storage. This would have broad implications not only for electric
vehicles but for other types of energy storage, a key requirement for making
the switch to wind and solar power on a global scale.
“This is a new way of doing battery development,” said Patrick Herring,
coauthor of the study and a scientist at the Toyota Research Institute.
“Having data that you can share among a large number of people in aca-
demia and industry, and that is automatically analyzed, enables much faster
innovation.”
The study’s machine learning and data collection system will be made
available for future battery scientists to freely use, Herring added. By using
this system to optimize other parts of the process with machine learning,
battery development—​and the arrival of newer, better technologies—​could
accelerate by an order of magnitude or more, he said.
60 Machine Learning Algorithms and Applications in Engineering

The potential of the study’s method extends even beyond the world of
batteries, Ermon said. Other big data testing problems, from drug devel-
opment to optimizing the performance of X-​rays and lasers, could also be
revolutionized by the use of machine learning optimization. And ultimately,
he said, it could even help to optimize one of the most fundamental processes
of all.
“The bigger hope is to help the process of scientific discovery itself,” Ermon
said. “We’re asking: Can we design these methods to come up with hypoth-
eses automatically? Can they help us extract knowledge that humans could
not? As we get better and better algorithms, we hope the whole scientific dis-
covery process may drastically speed up.”

4.1.5 A Review on AI-​based Predictive Battery Management


System for E-​Mobility
With the advancement in digitalization and availability of reliable sources of
information that provide credible data, Artificial Intelligence (AI) has emerged
to solve complex computational real-​life problems, which were challenging
earlier. The artificial neural networks (ANNs) play a very effective role in
digital signal processing. However, ANNs need rigorous main processors and
high memory bandwidth, and hence cannot provide expected levels of per-
formance. As a result, hardware accelerators such as graphic processing units
(GPUs), field programmable gate arrays (FPGAs), and application-​specific
integrated circuits (ASICs) have been used for improving overall performance
of AI-​based applications. FPGAs are widely used for AI implementation as
FPGAs have features like high-​speed acceleration and low power consump-
tion, which cannot be done using central processors and GPUs. FPGAs are
also a reprogrammable unlike central processors, GPU, and ASIC. In electric-​
powered vehicles (E-​Mobility), battery management systems (BMS) perform
different operations for better use of energy stored in lithium-​ion batteries
(LiBs). The LiBs are a nonlinear electrochemical system, which is very com-
plex and time variant in nature. Because of this nature, estimation of states
like state of charge (SoC), state of health (SoH), and remaining useful life
(RUL) is very difficult. This has motivated researchers to design and develop
different algorithms that will address the challenges of LiB state estimations.
This chapter intends to review AI-​based data-​driven approaches and hard-
ware accelerators to predict the SoC, SoH, and RUL of the LiBs. The goal is
to choose an appropriate algorithm to develop an advanced AI-​based BMS
that can precisely indicate the LiB states, which will be useful in e-​mobility.
An overview of the state-​ of-​
the-​
art on intelligent battery management
systems for electric and hybrid electric vehicles is also provided. The focus
is on mathematical principles, methods, and practical implementations. The
intelligent battery management systems aim at lengthening the lifetime of
the battery pack and enhancing the safety of drivers of electric and hybrid
Machine Learning Techniques Enabled Electric Vehicle 61

electric vehicles. Three major research topics are covered in the chapter, state
of charge (SoC), state of health (SoH) of the battery pack, and the remaining
driving range estimation.

4.2 Reverse Engineering with AI for Electric Power Steering


A prominent mystery of automobile industry scenario is that the vehicle
creators are more curious and covetous of knowing each and every minus-
cule insight concerning their rival’s vehicles. To discover the subtleties gen-
erally includes the study of start of art of the product and the concept of
dismantling a whole vehicle, piece by piece. The vehicle creators either have
their own inside groups who purchase a contender’s vehicle and dismantle
it or pay an external organization to do likewise. The reverse engineering
of vehicles is a concept that is carried out by some companies who well
specialized in it. They take pleasure in purchasing the most recent model of
any vehicle and fastidiously dismantling it like so numerous Lego blocks.
Apart from the delight, the reverse engineering concept can put forth some
great bucks by their attempts. They will offer solutions concerning their rival’s
vehicles to other vehicle creators. A full reported document of each thing that
went into the vehicle, alongside valuable added perspectives like how much
every segment probably costs. Moreover, they can assess the amount it costs
to accumulate the parts and give car producer knowledge into what sort of
gathering exertion their opposition is utilizing.
Manufacturers need to pay more attention to get entire knowledge about
their own vehicles and would definitely know how they accumulate their
own vehicles and what it costs. However, it tends to be extremely convenient
to see the evaluations made by the outsider. The customer may be high or
low, or in any case not all that demanding. It is additionally helpful to under-
stand what they are telling the rivals. In addition, it would then be able to
have them do a correlation for the parts and expenses related with the rivals
versus the vehicles.
A reverse engineering could convey whatever the client needed to think
about the vehicle, even the weight, size, and cost of each part along with the
ability of that part. The data about the producer of the part and piece of the
part is, for example, rates of metal versus different components could also be
given. Everyone discovers overpaying for their own carburetor and realizes
that to reduce costs and ultimately reduce the price, they charge for their car
that ought to switch. There’s an additional contort to this figuring out exer-
tion. In principle, in the event of being cautious to dismantle a vehicle, like-
wise examine the parts into a CAD framework and basically figuring out the
whole plan of the vehicle. With the present sophisticated CAD frameworks,
62 Machine Learning Algorithms and Applications in Engineering

it would then be able to see the vehicle from any point and jump into and
zoom all through the entire plan of the vehicle. Some truly modern CAD
frameworks will permit to siphon the plan into a test system program. This
could permit to conceivably go about like the vehicle exists and perceive how
it runs.
At the Cybernetic AI Self-​ Driving Car Institute, they are creating AI
frameworks for self-​ driving vehicles, which reveal an unmistakable fas-
cination that a similar industry-​wide quest for figuring out of customary
vehicles is presently in progress for AI self-​driving vehicles as well. The five
key stages to the preparing parts of an AI self-​driving vehicle are sensor infor-
mation assortment and translation, sensor combination, virtual world model
refreshing, AI activity plan refreshing, and car orders controls issuance. The
most noticeable actual parts of an AI self-​driving vehicle are the sensors. There
are a huge number of sensors on an AI self-​driving vehicle, including radar
sensors, sonic sensors, cameras, light detection and ranging (LIDAR), inertial
measurement units (IMUs), and different sensors. Every car creator and tech
firm is definitely keen on knowing which sensors different organizations are
utilizing in their AI self-​driving vehicles. It’s a fairly free-​for-​good now in
that there is no predominant provider fundamentally. To be sure, the different
organizations that make sensors appropriate for AI self-​driving vehicles are
in a savage battle about attempting to pick up gain traction for their sensors.

4.3 Artificial Intelligence Technology to Enhance


EV Production and Support the Deployment
of Electric Vehicles
The new AI innovation, which has been made via Artificial Intelligence,
is intended for modern scale EV assembling to improve the development
and assembling of batteries. EV battery packs are made out of an enormous
number of individual battery cells that are held and electronically associated
by various welded joints. High electrical resistance, because of helpless joint
quality, can make energy misfortune and warmth age, which can cause a
potential well-​being issue and furthermore lessens the productivity of the
battery. The AI innovation consequently evaluates the surface imperfections
and the electrical resistance of each joint before conclusive get together,
accordingly saving the maker’s time and expenses while additionally guar-
anteeing the well-​being of the battery.
Along with the improvement of smart grids, the wide appropriation of elec-
tric vehicles (EVs) is viewed as a motivation to the decrease of CO2 outflows
and more intelligent transportation systems. Specifically, EVs expand the
grid with the capacity to store energy at certain point in the network and give
Machine Learning Techniques Enabled Electric Vehicle 63

it back at others and, accordingly, help to improve the utilization of energy


from irregular environmentally friendly power sources and let clients refill
their vehicles in an assortment of areas. Nonetheless, various challenges
should be tended to if such advantages are to be accomplished. From one
viewpoint, given their restricted reach and costs associated with charging
EV batteries, it is essential to plan calculations and methodology that will
limit costs and, simultaneously, avoid clients being standard. Then again,
groups of EVs should be coordinated so as to stay away from peak on the
grid that may bring about high electric tariff costs and overburden local dis-
tribution grids.

4.4 Artificial Intelligence Used to Supercharge


Battery Development for Electric Vehicles
Battery operation can represent the deciding moment for the electric vehicle
experience, from driving reach to charging time to the lifetime of the vehicle.
Presently, artificial intelligence has made dreams in energizing an electric
vehicles in the time it takes to stop at a gas station a more likely reality and
could help to improve other aspects of battery technology. For quite a long
time, propels in electric vehicle batteries have been restricted by a significant
bottleneck—​assessment times.
Designing ultrafast charging batteries is a significant test, essentially in
light of the fact that it is hard to make them last. The intensity of the quicker
charge puts more prominent strain on the battery, which regularly makes it
fail early. To prevent this harm to the battery pack, a part that represents an
enormous piece of an electric vehicle’s all out cost, battery engineers should
test a comprehensive arrangement of charging techniques to locate the ones
that work best. The new exploration tried to upgrade this cycle. At the start,
the group saw that quick charging enhancement added up to numerous
experimentation tests, something that is inefficient for people, yet the ideal
issue for a machine. Machine learning is experimentation, however in a more
intelligent way and PCs are much better than us at sorting out when to inves-
tigate, attempt new and various methodologies and when to adventure, or
focus in, on the most encouraging ones. The group utilized this capacity for
their potential benefit in two key manners.
To start with, they utilized it to lessen the time per cycling test. In a past
report, the scientists found that as opposed to charging and energizing each
battery until it failed, the standard method of testing a battery’s lifetime. They
could predict how long a battery would last after just its initial 100 charging
cycles. This is on the grounds that the AI framework, subsequent to being
prepared on a couple of batteries cycled to disappointment, could discover
64 Machine Learning Algorithms and Applications in Engineering

designs in the early information that foretold how long a battery would last.
Second, machine learning diminished the quantity of strategies they needed
to test. Rather than testing each conceivable charging technique similarly, or
depending on instinct, the PC gained from its encounters to rapidly locate the
best conventions to test.
By testing less strategy for less cycle, the examination’s creators imme-
diately found an optimal ultrafast charging convention for their battery.
Notwithstanding significantly accelerating the testing cycle, the PC’s answer
was likewise better and substantially more uncommon than what a battery
researcher would almost certainly have conceived. It gave us this shockingly
straightforward charging convention something we didn’t expect. Rather
than charging at the most elevated current toward the start of the charge,
the calculation’s answer utilizes the most noteworthy current in the charge.
That is the distinction between a human and a machine: The machine isn’t
one-​sided by human instinct, which is ground-​breaking yet some of time
deceiving.
The specialists said their methodology could quicken virtually every bit of
the battery advancement pipeline: from planning the science of a battery to
deciding its size and shape, to finding better frameworks for assembling and
capacity. This would have wide ramifications for electric vehicles as well as for
different sorts of energy stockpiling, a critical necessity for doing the change
to wind and sunlight-​based force on a worldwide scale. The investigation’s
AI and information assortment framework will be made accessible for future
battery researchers to uninhibitedly utilize, Herring added. By utilizing this
framework to improve different pieces of the cycle with AI, battery advance-
ment and the appearance of fresher, better innovations could quicken by
a significant degree or more. The capability of the investigation’s strategy
broadens even past the universe of batteries.
Advances in electric vehicle batteries have regularly been restricted by
the colossal bottleneck of assessment times. At each phase of the battery
advancement measure, new advances should be tried for quite a long time or
up to years to decide how long they will last. Presently, Stanford University
analysts have built up an AI-​based strategy that cuts these testing times by
98 percent. The specialists tried the strategy on battery charge speed, how-
ever state that it tends to be applied to numerous different pieces of battery
improvement. The advancement strategy could make the fantasies about
energizing an electric battery in the time it takes to stop at a service station
a more probable reality. Artificial Intelligence is ready to add some juice to
electric vehicles by accelerating upgrades in battery innovation.
Specialists are drawing nearer to conveying batteries that highlight
upgrades profoundly looked for by architects and advertisers of elec-
tric vehicles: batteries that are more secure, revive quicker, and are more
practical than the age of lithium-​ion batteries now being used. Within five
years, specialists state, electric vehicles will arrive at value equality with
Machine Learning Techniques Enabled Electric Vehicle 65

conventional burning motor vehicles, thanks to some extent to the job AI is


playing. The battery makes up 25 percent of an EV’s complete expense, and
half of that battery cost is materials. Utilizing AI, battery analysts can accel-
erate the investigation and testing of a tremendous universe of new material
definitions that could make up a battery. The outcome is that progresses
in innovative work are going on at a speed up till now incomprehensible.
Disclosures that once would have taken a long time to accomplish now are
conceivable, now and again, very quickly.

4.5 How AI Helps Build Better Batteries


Working on quicker charging, more secure, and all the more impressive
electric vehicles, battery scientists are utilizing AI to accelerate the quest
for better materials and the testing cycle. Beneath, a glance at how AI could
improve lithium-​ion batteries. To comprehend the progressions occurring,
it assists with taking a gander at how ebb and flow battery research is
directed, and at lithium-​ion batteries, a battery type that packs a great deal
of energy into a little bundle, and that controlled the unrest in purchaser
gadgets and EVs.
Lithium-​ion batteries have three fundamental segments, yet each is
made out of numerous materials. This makes a practically boundless
universe of conceivable substitution materials when researchers need to
dabble to improve value, well-​being, or execution. Prior to AI, a quest for
new materials was finished with conventional measurable examination,
which was restricted by processing capacity and a specialist’s capacity to
gather data from the information. Man-​made intelligence, or AI, changes
the entirety of that. Quick PCs prepared to perceive explicit examples are
filtering through immense measures of information put away in the cloud,
taking advantage of materials’ information bases, analysis of results, and
long periods of logical writing, in a powerful hunt to recognize which
battery sciences are probably going to beat and which are probably going
to come up short.
At its examination lab in Almaden, California, International Business
Machines Corp. (IBM), has 1.94 percent researchers utilizing AI to quicken
improvement of a lithium-​ion battery that is quick charging and liberated
from hefty metals. Instead of utilizing nickel and cobalt, which are costly,
in restricted stock and hard to reuse, IBM says it has found an approach to
utilize an iodine-​based material that can be sourced from brackish waters
found in saltwater, conceivably making this battery a lot less expensive to
create than the present lithium-​ion details. One late arrangement of potential
materials for use in the electrolyte, a substance inside a battery that helps with
66 Machine Learning Algorithms and Applications in Engineering

the age of the electrical flow and had around 20,000 likely mixes. Utilizing
customary computational techniques to screen those competitors might have
taken five years. AI assessed them in nine days. The AI is currently getting
substantially more important as we change the solvents and electrolytes to
improve regarding limit and life cycle, since we have had more opportunity
to prepare it.
Notwithstanding assisting with materials research, AI decreases the
time it takes to test batteries. Previously, streamlining new battery-​cell
plans was a cycle that regularly took long stretches of charging and
releasing the batteries a great many occasions. Presently, AI’s capacity
to rapidly investigate tremendous measures of information gathered
during battery testing permits the researchers to make forecasts about
execution a lot quicker, decreasing the quantity of tests that should be
run. The anode stores lithium and deliveries lithium-​ion (lithium less an
electron) when the battery is releasing. The separator permits lithium
ion to go through while electrons are compelled to travel independently
delivering an electric flow. What’s more, the AI empowers specialists to
recognize incorporated natural mixes that could help the anode’s ability
to hold lithium ion. The cathode works when the battery is revived, and
the lithium ion and electrons venture out back to the anode in different
ways. What’s more, with AI’s assistance, scientists are building up an
iodine-​based option in contrast to cobalt, a costly, hard to reuse hefty
metal utilized in most lithium-​ion batteries. The electrolyte, normally a
blend of salts and solvents, encourages the development of the lithium
ion. With AI scientists can test new electrolyte plans quicker, assisting
them with distinguishing mixes with expanded voltage and higher blaze
focuses. Quicker charging commonly harms the separator, lessening the
battery’s lifetime and possibly prompting fires. Computer-​based intel-
ligence assists analysts with finding the sweet spot adjusting charging
speeds, charging flows, charging recurrence, and battery lifetime.
Researchers and analysts around the globe are progressively going to
Artificial Intelligence (AI) and alleged robo-​scientists to assist them with
finding everything from new anti-​infection agents and medications, through
to new shower on sun-​powered board materials and immunizations, so it
shouldn’t come as a very remarkable amazement that AI is presently being
utilized to help grow new battery tech. “In battery testing, you need to attempt
countless things, on the grounds that the exhibition you get will change def-
initely,” said Ermon, an associate educator of software engineering at MIT,
alluding to how researchers customarily chase for new battery forward leaps,
and who drove the new undertaking to utilize AI to assist them with cre-
ating promising batteries for electric vehicles. “With AI, we’re ready to rap-
idly recognize the most encouraging methodologies and cut out a great deal
of pointless trials.” The examination, distributed by Nature, was essential
for a bigger cooperation among researchers from Stanford University, MIT,
Machine Learning Techniques Enabled Electric Vehicle 67

and the Toyota Research Institute that spans basic scholastic exploration and
genuine industry applications, and their objective: to locate the best strategy
for charging an EV battery in a short time that amplifies the battery’s general
lifetime.

4.6 Machine Learning Supercharges Battery Development


The analysts composed an AI program that, based on few charging cycles,
anticipated how batteries would react to various charging approaches. The
product likewise chose continuously what charging ways to deal with center
around or disregard. By lessening both the length and number of preliminaries,
the analysts cut the testing cycle from very nearly two years to only 16 days.

4.7 AI-​Based Predictive Battery Management System


for E-​Mobility
This chapter additionally features the specialized difficulties and arising
innovations for the improvement of effectiveness, unwavering quality, and
security of EVs in the coming stages as another commitment. The advantages of
electric vehicles rise up out of these vehicles’ capacity of continuing their energy
requests through electric network as opposed to petroleum product utilization.
And also environmental examinations have demonstrated that electric drive
(E-​drive) offers the most elevated eco-​friendliness and subsequently the least
discharge of greenhouse gases. The worldwide stores of diesel, petroleum, and
other nonrenewable energy sources are diminishing quickly because of their
broad use in transportation activity. The conventional power sources produce
huge loads of CO2 yearly, which have hurtful ramifications for the climate, for
example, ozone-​depleting substance discharges and a dangerous atmospheric
deviation. Moreover, the expenses of these powers are expanding dramatically,
so there is a requirement for an auxiliary fuel hotspot for transportation, for
example, electric vehicles (EVs), new energy vehicles (NEVs), plug-​in hybrid
vehicle (PHEVs), battery electric vehicles (BEVs), and fuel cell electric vehicles
(FCEVs). As of late, battery-​powered batteries have pulled in significant con-
sideration inferable from their appeal in EVs, HEVs, and PHEVs. With the util-
ization of sustainable power, these transportation sources can decrease their
GHGE by up to 40 percent. The elective fuel sources, for example, wave, wind,
flowing, and sunlight-​based, are episodic, so these energy assets additionally
require an energy storage systems (ESS) to keep a smooth and dependable
inventory.
68 Machine Learning Algorithms and Applications in Engineering

The battery management systems (BMS) is a basic part of electric and


hybrid electric vehicles. The motivation behind the BMS is to ensure protected
and dependable battery operation. To keep up the security and unwavering
quality of the battery, state observing and assessment, charge control, and cell
adjusting are functionalities that have been actualized in BMS. As an elec-
trochemical item, a battery operates contrastingly under various operational
and natural conditions. The vulnerability of a battery’s exhibition represents
a test to the execution of these capacities.
With the evolution in digitalization and accessibility of reliable sources
of data that give sensible information, Artificial Intelligence (AI) has been
developed to tackle a complex computational genuine issue, which was
inspiring before. The ANNs assume an extremely compelling job in digital
signal processing. ANNs need thorough principal processors and high
memory data transfer capacity, and thus can’t give anticipated degrees of
execution. The graphic processing units (GPUs), field programmable gate
arrays (FPGAs), and application-​specific integrated circuits (ASICs) have
been utilized for improving the overall performance of AI-​based applications.
FPGAs are generally utilized for AI usage as FPGAs have highlights like
rapid quickening and low power utilization, which is impossible when util-
izing local processors and GPUs. FPGAs are additionally a reprogrammable
dissimilar to local processors, GPU, and ASIC. In electric-​powered vehicles
(e-​mobility), BMS perform various tasks for better utilization of energy stored
in lithium-​ion batteries (LiBs).
The LiBs are a nonlinear electrochemical system, which is extremely
complicated and time variant in nature. Due to this nature, assessment
of states like state of charge (SoC), state of health (SoH), and remaining
useful life (RUL) is challenging. This has inspired scientists to plan and
create various calculations that will address the difficulties of LiB state
assessments. Battery capacity should be intended to permit EV drivers
arrive at their objective endpoint while evading unnecessary stops to ener-
gize their vehicles. Nonetheless, this extra battery capacity would affect the
vehicle’s space, weight, and cost. Considering these restrictions, the coord-
ination of EVs with the vision of intelligent transportation systems (ITS) is
integrated.
This section begins by placing the plan of EVs into a more extensive view-
point by proposing a Predictive Intelligent Battery Management System
(PIBMS), which will upgrade the general exhibition of EVs ,including
energy utilization and outflows utilizing the ITS foundation. It spreads out
the plan establishment for the future execution of an interconnected EV
furnished with PIBMS, which further adds to the improvement of energy
proficiency and diminished emanations. Energy stockpiling framework
(ESS) innovation is as yet the logjam for the electric vehicle (EV) industry.
Lithium-​ion (Li-​ion) batteries have pulled in impressive consideration in
the EV business attributable to their high energy density, life expectancy,
Machine Learning Techniques Enabled Electric Vehicle 69

nominal voltage, power density, and cost. In EVs, a smart battery manage-
ment system (BMS) is one of the fundamental parts; it gauges the conditions
of battery precisely, yet in addition guarantees safe activity and drags out
the battery life. The exact assessment of the condition of charge (SOC) of a
Li-​ion battery is an extremely testing task in light of the fact that the Li-​ion
battery is an exceptionally time variation, nonlinear, and complex electro-
chemical framework.
This section clarifies the functions of a Li-​ion battery, gives the fundamental
highlights of a smart BMS, and thoroughly surveys its SOC assessment
techniques. These SOC assessment techniques have been ordered into
four fundamental classes relying upon their inclination. A critical clarifica-
tion, including their benefits, restrictions, and their assessment errors from
different examinations, is given. A few proposals relying upon the advance-
ment of innovation are recommended to improve the online assessment.
This section tends to the worries for the current BMSs. State assessment of a
battery, including condition of charge, condition of well-​being, and condition
of life, is a basic undertaking for a BMS. Through inspecting the most recent
systems for the state assessment of batteries, the future difficulties for BMSs
are introduced and potential arrangements are proposed too.

4.8 Uses of Artificial Intelligence for Electric Vehicle


Control Applications
The electric vehicle (EV) is emerging as the best-​in-​class innovation vehicle
tending to the consistently squeezing energy and environmental concerns. To
diminish the dependence of oil and natural contamination, the development
and integration of electric vehicles has been quickened in numerous nations.
The execution of EVs, particularly battery electric vehicles, is viewed as an
answer to the energy emergency and natural issues. This chapter provides
significance to specialized improvement of EVs and arising advances for
their future application. Key advances with respect to batteries, charging
innovation, electric engines and control, and charging foundation of EVs
have been summed up.
This section presents a novel speed control plan of electric vehicle (EV)
to improve the comportment and strength under the condition of various
road requirements. Parameters that direct the working of PI regulator are
progressively changed with the help of fuzzy intelligence control. All things
considered, electric vehicles (EVs), including full cell and hybrid vehicles,
have been developed quickly as an answer for the energy and ecological
issues. Driven EVs are controlled by electric engines through transmission
and differential gears, while straightforwardly determined vehicles are
70 Machine Learning Algorithms and Applications in Engineering

pushed by in-​wheel or, basically, wheel engines. The fundamental vehicle


configurations of this exploration have two straightforwardly determined
wheel engines introduced and worked inside the driving wheels on a pure
EV. These wheel engines can be controlled autonomously and have so quick
and exact reaction to the order that the vehicle skeleton control or movement
control turns out to be more steady and vigorous, contrasted with, by impli-
cation, driven EVs. Like most exploration on the torque distribution control of
the wheel engine, the wheel engines provided a powerful ideal tractive power
conveyance control for an EV driven by four-​wheel engines, in this manner
improving vehicle handling and stability. The analysis has demonstrated
that EV control techniques, for example, PI control can perform optimally
over the full range of operations and disturbances and it is viable with con-
sistent vehicle torque. Moreover, the nonlinear vehicle torque is not fixed
and changes arbitrarily. Anyway, EV with regular PI control might not have
agreeable execution in such quick fluctuating conditions, the system perform-
ance deteriorates. Also, it is hard to choose appropriate control boundaries
Kp and Ki to accomplish acceptable compensation results while keeping up
the stability of EV traction, because of the highly complex, nonlinear nature
of controlled systems. These are two of the significant disadvantages of the
PI control. To beat these challenges, adaptive PI regulator by fuzzy control
has been applied both in fixed and under streets limitations and appears to
improve the general presentation of EV.
The direct torque control (DTC) methodology is one sort of superior per-
formance innovations for AC motors, because of its simple structure and
capacity to accomplish quick reaction of flux and torque, due to which it
has attracted interest in the recent years. DTC-​SVM with PI regulator direct
torque control without hysteresis band can successfully diminish the torque
ripples; however, its system robustness will hide its improvement. DTC-​SVM
strategy can improve the system robustness, clearly diminish the torque and
flux ripple, and adequately improve the dynamical responses. The DC–​DC
converter is used with a control methodology to guarantee the energy need
for the EV and the drive system. The objective of this section is to add to
understanding the intelligent fuzzy PI regulator for utility EV tow rear deter-
mining wheel applied direct torque control-​based space vector balance under
a few situations.
The commercialization of electric vehicles/​hybrid electric vehicles (EV/​
HEV) is up to now being tested with respect to performance and cost. The
Cost decline should be investigated as an ideal plan. Other featured research
includes green drive plans, which incorporate the integration of possible
supportable renewable energy sources to make eco-​obliging green vehicles,
similar to the Internet of Things (IoT)-​based methods for EV/​HEVs. Electric
vehicle research incorporates multi-​ disciplinary capacity from electrical
equipment.
Machine Learning Techniques Enabled Electric Vehicle 71

4.9 Conclusion
There is an increasing demand for scalable and autonomous management
systems. We have proposed AI-​ based EV, which would eliminate fossil
fuel (petrol and diesel) consumption. A universal approach to performing
the reverse engineering of electric power steering (EPS) for the purpose of
external control is also presented, designed, and executed. The primary
objective of the related study was to solve the problem associated with the
precise prediction of the dynamic trajectory of an autonomous vehicle, which
was presented and designed. This task has been accomplished by deriving
a new equation for determining the lateral tire forces and adjusting some of
the vehicle parameters under road test conditions. The expert systems were
made more flexible and effective for the present application by the introduc-
tion of hybrid artificial intelligence with logical reasoning. The innovation
offers a solution to the major problem of liability in the event of an autono-
mous transport vehicle being involved in a collision.
5
A Comparative Analysis of Established
Techniques and Their Applications
in the Field of Gesture Detection

Muskan Jindal, Eshan Bajal, and Shilpi Sharma


Department of Computer Science and Engineering, Amity University,
Uttar Pradesh, India
E-​mail: [email protected], [email protected],
[email protected]

CONTENTS
5.1 Introduction................................................................................................... 74
5.2 Challenges and Areas of Improvements................................................... 76
5.2.1 Image Acquisition Challenges........................................................ 76
5.2.2 Gesture Tracking, Segmentation and Identification
Challenge........................................................................................... 76
5.2.3 Feature Extraction Challenges in Gesture Detection and
Identification..................................................................................... 77
5.2.4 Limitations and Challenges in Gesture Gradation or
Categorization................................................................................... 78
5.2.5 Limitations of End-User Gesture Analysis
Customization................................................................................... 79
5.3 Related Fields and Special Mentions......................................................... 79
5.3.1 Collaborative Learning of Gesture Recognition and
3D Hand Pose Estimation with Multi-Order
Feature Analysis............................................................................... 79
5.3.2 Gesture Analysis and Organizational Research: 
The Development and Application of a Protocol for
Naturalistic Settings......................................................................... 80
5.3.3 Gesture and Language Trajectories in Early
Development: An Overview from the Autism Spectrum
Disorder Perspective........................................................................ 80
5.3.4 Learning Individual Styles of Conversational Gesture............... 81

DOI: 10.1201/9781003104858-5 73
newgenprepdf

74 Machine Learning Algorithms and Applications in Engineering

5.3.5 Sleep Gesture Detection in Classroom Monitor System............. 81


5.3.6 Towards the Markerless and Automatic Analysis
of Kinematic Features: A Toolkit for Gesture and
Movement Research......................................................................... 82
5.4 Analysis of the Contemporary Literature Pertaining to Gesture
Detection........................................................................................................ 82
5.5 Conclusion..................................................................................................... 87

5.1 Introduction
In day-​to-​day mundane conversations and tasks, information reinforcement
and clarification play a cardinal role, using sign language as a medium to
do so. Sign language is an irreplaceable component of daily communication,
and it may have variegated forms like visual motions or hand gestures, but
it aims to convey the same message and fulfil the purpose of information
reinforcement. It includes use of multiple body parts including fingers and
hands, which are the popular ones, but it also incorporates facial gestures,
head, body and arms, which may be lesser known in mundane communi-
ties. For those with hearing, listening or visual disabilities, sign language is
the only means of communication and expression. But those with these dis-
abilities depend completely on sign language to interact with outer world
find it cumbersome to do so as many people are unaware of the complete
imbrications of sign language. Moreover, while use and knowledge of sign
language is common among communities with visual disabilities, it is not
very widespread among communities with hearing disabilities. This poses
a challenge for these communities to interact or convey their message to the
outer world as they have relied completely on sign language, which con-
tinues to be an unsolved issue till today.
While most of the sign language includes upper half of body [1]‌, some
sign language methods use variegated shapes and multiple changes in the
complete body [2]. Gestures most commonly used are hand gestures, which
can be classified into multiple types, namely, conversational gestures, con-
trolling gestures, manipulative gestures and communicative gestures [3].
Due to the highly structured and organized attributes of sign language, the
field of computer vision and its various algorithms suits the same [4]. The
study here compiles contemporary expertise on the detection of gestures
implementing supervised learning and image segregation algorithms,
along with their associated mannerisms while proposing modifications for
inclusivity of a wider audience, with a prominent emphasis on people with
communication disabilities. Furthermore, the work done in this chapter may
culminate in the development of real-​world applications such as determining
A Comparative Analysis of Techniques in Gesture Detection 75

the attentivity of students in real time, which is an indispensable asset in


the post-​COVID world of online interactions. Implementation of computer
vision algorithm for gesture analysis is suitable for the complete process of
identifying, monitoring and tracking various gestures performed while com-
municating in sign language because the complexity and structured nature
of the process can be best interpreted by the nature of computer vision meth-
odologies. Though the field seems inchoate and nascent, some of the earliest
research literature was published in 1933, where multiple gesture analysis
techniques were utilized to acclimate to available speech and handwriting
material for further analysis and identification. Few of the same include
technique implemented by Pentland and Darrell, which aided in the iden-
tification of multiple dynamic structures by using an optimized dynamic
time warping (DTW). Eventually [5], hidden Markov models (HMMs)
were implemented to comprehend and analyse various subtle attributes or
properties to eventually categorise the shape and trajectory of various sign
languages. Further, another literature elucidated the accuracy of the same
by collating 262 signatures from variegated signers to obtain and calculate
to an accuracy of around 94 percent, proving the method viable. The reason
behind deliberating signs from multiple signers was that it was observed that
considering signs from a particular individual though increased accuracy on
training set but decreased the actual accuracy of model [6]. Later building on
the concepts of HMMs, another pair of prolific researchers, namely, Vogler
and Metaxas [7] elucidated the multiple limitations of HMMs while using the
dependent model. [8] Ascension Technologies Flock of Birds devices collated
three-​dimensional translation and rotation data of the sign by the aid of mul-
tiple authors to facilitate researchers to further analyse the area of research.
Another very popular and well-​developed area of gesture analysis, specif-
ically hand gesture hand detection, is human–​computer interaction, via a UI/​
UX based interface facilitated by a processing unit, where data is collected
via visual or sensory devices like camera by video or imageries [9]‌. Multiple
receptor devices like the one that can be worn for data acquisition include
data gloves [10] that facilitate accurate and abundant access to data for ges-
ture analysis or detection. Other sensory receptor devices, including camera,
solve the issues like moment of hand gestures, which are the limitation of
data gloves but create their own indigenous problems like camera orienta-
tion, background lights and foreground disturbance. Thus, data acquisition
remains a cumbersome step in the gesture detection process but can be eased
by multiple devices and frameworks like Microsoft Kinect [11], Leap motion
controller [12] and ASUS Xtion [10]. The next step is to effectively learn the
acquired data to comprehend and identify various patterns, similarities and
deduce conclusions from the same. This task of learning from raw data can
be performed easily by the application of variegated learning algorithms like
deep learning, convolutional neural network (CNN), adapted convolutional
neural network (ADCNN) and recurrent neural network (RNN) [13,14].
76 Machine Learning Algorithms and Applications in Engineering

5.2 Challenges and Areas of Improvements


While there are multiple steps of gesture analysis as described by multiple
authors, some gradation steps create more challenges than other methods.
After discerning through multiple research articles and literature published
till date, given below are the selected methods that create more challenges in
the process of gesture analysis and recognition.

5.2.1 Image Acquisition Challenges


Capturing or acquiring images is done via multiple sources by the aid of
camera or other image-​capturing devices. This is the primary step of any
gesture analysis or recognition framework and algorithm. When images
are captured from an analysis perspective, for example those captured by
a person or researcher, physical aspects like angle of picture and back-
ground colour contrast are taken care of, but when this task is performed
by an automated machine, image capturing might not be done effect-
ively [15]. Issues like image capturing frame and background colour con-
trast can create further issues in steps like gesture detection and unique
gesture identification. While hand gesture detection frameworks have
become robust and effective, computer vision is still an abstruse con-
cept to people and image-​capturing cameras are placed properly by the
individual using the same. So, the placement of camera creates issues in
later phases of gesture analysis, detection and identification [16]. Other
issues in image capturing include loss of depth, varying speed of the ges-
ture performer, inability to identify the difference between two gestures
and spatial or temporal inconsistencies [17]. While the above-​mentioned
constraints are specific to a scenario, there are some very mundane issues
that occur during the same including complicated or busy background,
loss of focus due to busy backdrop, real-​time monitoring constraints and
others. The image capturing object, that is, camera and its specifications,
also plays a vital role in image acquisition and its attributes such as colour,
quality, resolution and number of frames per unit time or frame rate. 3D
image capturing amplifies these constraints as more accurate and high-​
resolution camera is required here.

5.2.2 Gesture Tracking, Segmentation and Identification Challenge


Gestures have a certain degree of rotation, that is, the degree or directional
radius where they move freely; the higher is their degree of rotation, higher
is the complexity in gesture analysis and identification. While this appears
A Comparative Analysis of Techniques in Gesture Detection 77

as the major challenge especially in hand gestures, other limitations include


inconsistent speed of gestures [18], inability to distinguish between start
and end of gestures and variation on skin texture or complexion. Other ges-
ture image segmentation challenges include inconsistencies on background
like variation in illumination, busy or moving background, and continuous
differences in colour patterns of the background [19].

5.2.3 Feature Extraction Challenges in Gesture Detection


and Identification
There are multiple features that are extracted via variegated approaches
that face varied challenges and graph-​based features when extracted face
limitations due to complex background or inconsistent illumination [19].
Mixing of gestures or mixing of features also creates several limitations in the
same. While the degree of rotation is well accustomed with various gesture
detection algorithms, abrupt gesture or out-​ of-​
defined boundary gestures
create issues. This gradation of limitation also incorporates the challenges
and issues that occur during the identification of continuous gestures as the
detector deliberates them as one continuous gesture instead of multiple unique
gestures. A solution to this issue of continuous gestures was proposed by intro-
ducing automatic system-​based pauses since these pauses may or may not
coincide with the actual pauses taken by the person. Moreover, many times
due to personal habits, some people perform gestures by taking few or no
pauses in between while others perform gestures with an optimal number of
pauses [20]. These discrepancies also create inconsistences in the identifica-
tion of gestures when gestures are performed by humans, but if performed by
artificial devices, such issues of inconsistencies tend to be avoided [21]. Apart
from the above, another type of gestures, namely ‘conversational gestures’,
also have similar issues on grounds of gesture identification and segmentation
as they have multiple variations based the personal way of using conversa-
tional gestures, variation in dialect as different dialects and regional influences
create variations in speed, frequency and type of conversational gestures used.
Real-​time conversational gestures also differ widely based on the environment
like public gestures and private gestures, which can create further discrepan-
cies in both gesture identification and segmentation. Additionally, emotional
influence on gestures is a factor that creates inconsistencies in the same [22,23].
There are categories of gestures that create problems of conscious and uncon-
scious gestures that humans perform during mundane conversations. While
this can be considered as a subcategory to conversational gestures, they can
be deliberated more effectively when considered individually [24]. Thus, real-​
time gesture identification, analysis and segmentation is a cumbersome and
complex task, which can be improved by continuous training over a long
period of time.
78 Machine Learning Algorithms and Applications in Engineering

5.2.4 Limitations and Challenges in Gesture Gradation


or Categorization
A gesture-​detecting algorithm’s final gradation is to categorise the accumulated
gestures into multiple categories, namely, supervised and unsupervised.
Challenges occurring during this phase occur mainly during handling large
data points like higher complexity data points, categorising ambiguous data
into pre-​defined classifications, identification of appropriate data sets that sat-
isfy most of the data points, handling outlier data points and other limitations
like handling ambiguous gestures [25]. Multiple challenges occur while
implementing variegated classifiers and filters when classifying captured
gestures. When implementing classifiers like dynamic time wrapping (DTW)
classifier, paucity of huge number of data sets and vocabulary size of the same
create a limitation. For distance-​dependent algorithms like k-​nearest neigh-
bour, the large data set increases the computation complexity exponentially,
thereby increasing time and space complexity as well, depleting the overall
performance of the algorithm. Much like DTW algorithms, deep learning-​
based classification approaches also deplete the performance due to paucity
of data sets and also require huge training or validation set to provide
optimal performance. Primordial or classical methods of classification like
finite-​state machines have very rigid transform states and boundaries, which
is a palpable limitation of the same. Cluster-​based classifiers like k-​means
classification approach is predominantly dependent on the initial cluster and
its attributes, which create discrepancies later, and also it is influenced nega-
tively by the presence of classifiers. So choosing an appropriate value for ‘k’
is a cumbersome task. The similar issue of selecting optimal value is faced
by sister classifiers known as support vector machine (SVM), where selecting
appropriate kernel value becomes a daunting task. Moreover, SVM has other
limitations like its high computation complexity, which increase the expenses,
and being a binary classifier by nature, pairwise classification is required for
handling multiple class classification as illustrated in Figure 5.1 below.

FIGURE 5.1
Limitations and Challenges in Gesture Gradation or Categorization are elaborated pictorially.
A Comparative Analysis of Techniques in Gesture Detection 79

5.2.5 Limitations of End-​User Gesture Analysis Customization


This is a comparatively special and new development in the field of gesture
analysis, recognition and segmentation. It deliberates the scenario where the
end-​user requests for custom-​specific gestures to be incorporated and fur-
ther activities based on the same [26]. Wobbrock created a set of 27 novel
gestures along with their respective activities performed as reaction to those
gestures. Post that, many authors started adding custom gestures to their
work along with their subsequent reactions as a response [27]. The take-
away from these and many following research is to comprehend the fact that
instead of creating personal or specific user gestures, the development of ges-
ture that can be universally understood, that is, by a large number of people,
is the most viable way to go as it includes the concept of gesture creation.
While there are numerous studies that include custom gestures [28–​31], but
most of the mentioned studies do not provide inclusion of gesture creation.
Instead, personal gestures are either pre-​trained and included in the frame-
work, ready for the end-​user to use or provided to user to train of himself/​
herself. Multiple areas of application are provided for custom gestures like
password, specific actions, or just performing mundane tasks in general.

5.3 Related Fields and Special Mentions


Along with a major literature review about the broad techniques of gesture
analysis, this section deals with description and analysis of some latest state-​
of-​art literature in gesture analysis, while also focusing on their specific area
of applications.

5.3.1 Collaborative Learning of Gesture Recognition and 3D Hand


Pose Estimation with Multi-​Order Feature Analysis [32]
While gesture analysis and recognition is itself very abstruse task spe-
cially when implemented for feature analysis in hand gestures as various
challenges arise as elucidated above, approximating a 3D hand pose in real
time opens an altogether new avenue of challenges and problems. This latest
state-​of-​art research study performs 3D hand gesture approximation and rec-
ognition with multiple order feature recognition by implementing a novel
framework of a network exploring joint aware attributes that appear promin-
ently in all occurring features. It also partly addresses the issue of paucity of
data by deliberating the considered task’s annotation as a weak supervision.
Moreover, to perform both feature and hand pose recognition in 3D poses,
it leverages common structures and gestures throughout the hierarchical
architecture. This research is very well structured and written with astute
80 Machine Learning Algorithms and Applications in Engineering

recognition to performance and novelty. The framework proposed is well


elaborated and explained with appropriate architectural diagrams to assist
the same. Based on the results on the grounds of performance, accuracy and
space–​time complexity, the methodology is phenomenal.

5.3.2 Gesture Analysis and Organizational Research: The Development


and Application of a Protocol for Naturalistic Settings [33]
This one-​of-​a-​kind research study performing the very complex yet essen-
tial task of analysing, interpreting and deriving conclusion from gestures
people use in real time to gather insight about their behaviour, intentions
and motives. It includes analysing real-​time gestures of people while cap-
turing their gestures and body language via video input to get insights, for
example analysing the same from video input of presentations, organization
and leadership summits. While the presented work may not be able to pro-
vide specific details and exact connotation, it successfully elucidates some
general protocols and useful outputs when given different input videos and
images. It can make very astute observations about the subject of observa-
tion, domain of topic discussed in the input video, general intensity of the
environment and formal/​informal discussion domain. It classifies gestures
in an entrepreneur environment by using verbal or non-​verbal indicators.
After appropriate training, it succeeds in classifying entrepreneurship and
leadership talks into different topics by just analysing gestures of the speaker
and his/​her body language. Moreover, it can also recognize a particular
author by his/​her unique style of gestures, body and verbal communication
after appropriate training.

5.3.3 Gesture and Language Trajectories in Early Development: An


Overview from the Autism Spectrum Disorder Perspective [34]
This study describes a lesser known but substantial domain area of applica-
tion of gesture analysis and recognition, that is, autism spectrum disorder
(ASD) and its close relationship to gesture analysis. Gesture recognition and
body language analysis along with nonverbal communication parameters
play a cardinal role in early diagnosis of ASD and its regrowth of symptoms.
This aims to fulfil two primary purposes in the mentioned domain, namely,
verifying the extent of validity of notion that gesture analysis can aid in diag-
nosis of ASD, and comprehending the prominent differences between the
language loss trajectory of various types of autism in early ages of children
from various aspects like qualitative aspects, quantitative aspects and more.
By the end of this research, the author successfully verifies the validity of
the notion and concludes that consistent impairment or delay in response
gestures indicates early development of ASD. Additionally, after intensive
literature study, it concludes that teaching extra gestures or nonverbal means
of communication can not only help these children fulfil the need of social
A Comparative Analysis of Techniques in Gesture Detection 81

interaction but also build up their confidence to further aid in the treatment
of the disorder. After discerning differences in gesture analysis, the research
identifies multiple differences between trajectories of frequency of gestures
in different stages of ASD, as well differences based on the age of the patient,
tracking trajectory can also indicate the course, speed and effectiveness of
treatment, if given. Apart from the above, early diagnosis and detection can
also be related to the neurological paths that patients follow while their dis-
order regrows or slips in the dormant state. These insights can be substantial
in finding cure to their respective disorders.

5.3.4 Learning Individual Styles of Conversational Gesture [35]


While there are multiple research papers exploring conversational styles, non-
verbal gestures, body language and provide variegated models for predicting
the subsequent next gestures of their subject of observation after training
their model, this research provides ground-​breaking results in the domain
for multiple speakers. It deliberates speaker-​specific data sets of 144-​hour
video imageries analysing and training in a data-​driven technique, to label,
clean, pre-​process and annotate the data. Post this step, it uses a combination
of methods, namely, U-​Net architecture for speech to gesture conversion,
combination of regression models to predict subsequent motion of the object
of observation (speaker in the video) to get the predicted pose sequence, to
finally evaluating the obtained results. To further verify the robustness of the
proposed algorithm and methodology, Amazon Mechanical Turk is used to
conduct a real versus fallacious gesture analysis by providing input to the
same. Although the predicted sequence of gestures is respective to the object
of observation and are also influenced by the other inter-​personal styles
that are different for different speakers, this model succeeds in providing
high performance for variegated speakers and styles. By the end of study, it
provides real-​time gesture analysis both qualitatively and quantitatively for
data set with path-​breaking performance.

5.3.5 Sleep Gesture Detection in Classroom Monitor System [36]


Sleep gesture identification has always been a potential field of application or
extended application for gesture analysis but has been comparatively unex-
plored due to multiple challenges and its demand of being robust considering
it has wide real-​life application. This research overcomes the limitations of
sleep gesture identification like challenge of occlusion, variegated nature of
sleep gestures, misidentification of the same with other gestures similar in
nature and inability to implement the model in real-​life data. The presented
state-​of-​art research collates data from real-​life classes of Shanghai, pre-​
processes the same and implements optimized R-​FCN amalgamated with
feature pyramid and deformable convolution for gesture recognition. It
82 Machine Learning Algorithms and Applications in Engineering

overcomes specific challenges encountered with real data like recognizing


small sleep gestures and also partially identifying loss of attention span of
students occurring just before or after sleep gestures. This methodology not
only provides ground-​breaking performance with noisy video or digital
imageries as input.

5.3.6 Towards the Markerless and Automatic Analysis of Kinematic


Features: A Toolkit for Gesture and Movement Research [37]
Given the paucity of frameworks that perform robust motion detection, iden-
tification of spatial features for real-​life data is solved by the presented work.
The reason for this is the loss of robustness due to use of hard coding for the
same, inability to identify spatial features due to limitation like distance, one
at a time approach and limited processing speed. These challenges are very
astutely identified and handled by the proposed approach, which captures
frequently used or repeated feature of the object of observation, spatial and
kinetic attributes. It has wide application and uses due to its robust and
dynamic nature.

5.4 Analysis of the Contemporary Literature Pertaining


to Gesture Detection
In this section, papers that stand out from the multitude of literature
associated in any shape or form with the use of artificial network-​based
learning techniques for the benefit of humanity are considered. Although
the fecundity of the works of the literature mentioned cannot be assimilated
in such a small place, each piece of literature is explained in an elaborate
manner to provide astute reliable insight into the papers without delving into
each one of them individually.
Razieh et al. [38] performed a thorough survey on the current deep learning-​
based gesture detection frameworks available out there. The chapter starts by
listing all the available features used to determine a framework followed by
a brief introduction of the parameters. In the next section, all the material
collected is sorted based on the type of input for analysis namely, two-​
dimensional or three-​dimensional and the number of colour channels used.
Other parameters such as visual modality and the difference between static
and dynamic imagery has also been briefly introduced, followed by a chart
depiction of all the papers analyses, the corresponding datasets used and the
accuracy of the results in percentage or similarity and confidence metrics. This
comparison cumulates papers from 2001 to 2019 and includes data sets such
as NYU, ICVL, MSRA, LSP, FLIC, ITOP, STB and isoGD, among the notable
A Comparative Analysis of Techniques in Gesture Detection 83

ones. Multiple discussion sections deal with different aspects of working


with models of gesture analysis. The first section discusses the importance
of facial features along with suitable examples and elucidates its importance
in semantic analysis. This second discussion section deals with the different
models usable and their structural importance such as advantages of using
ANN over CNN, while the third section discusses the work related to the
hybrid approach using the combination of classical classifiers with the deep
learning-​based descriptors.
Mohammed, in his study [39], reviewed models with a strong focus on the
Arabic sign language and reviewed all the pieces of literature that delve into
it. Various learning frameworks such as CNN · RNN · MLP · LDA · HMM ·
ANN · SVM · KNN and SLR systems were explored with elaborations and
corresponding accuracies of each of them. One important factor to realize
is that unlike its American counterpart, the Arabic sign language systems
require both hands to convey some alphabets, which means more spatial
awareness is necessary on the parts of the models. Among the 45 papers
reviewed, Hartanto’s technique scored the lowest with accuracy of 62.8 per-
cent while others scored a mean of 91.2 percent. From these works, it has
been concluded that the deep learning-​based frameworks rival the analysing
capabilities of humans and the neural networks working built on multilayer
perceptron possess the highest performance to requirement ratio.
Another real-​time application of deep learning in the realm of hand gesture
detection was done by Murat et al. [40] on the Tensorflow framework with the
Kearas library. The proposed CNN model first converts the coloured image
into the YCbCr space to detect the hand using colours. Next, the convex hull
algorithm determines the position and location of the hand so that the image
can be resized and loaded for classification. These two steps are carried out
simultaneously after which a greyscale conversion precluded the feeding
into the machine in 30 ms loops. This system was highly capable, scoring
99.44 percent to 100 percent in the validation set. Along with the graphs and
images, the real-​time working of the system is elucidated.
Hamid et al. [41] created a wearable wrist-​mounted device to capture the
hand movement data for processing via a deep neural network. Using tri-
axial accelerometers and gyroscopes mounted on Inertia Measuring Devices
tied to a subject’s hands, the passive data during feeding was collected that
was further processed to analyse the actions performed. Using a two-​step
process of identifying intake frames and peak activity frames, the data of
100 individuals was evaluated. After accounting for discrepancies for gravity,
the data was processed by the four-​layered CNN. According to the results,
the system reached a peak accuracy of 88 percent with a varying range. The
major drawbacks of the paper come from the system flagging nondeterminate
actions like blowing air falsely. The presence of more pooling layers did not
drastically change performance neither did changes in the FLOP parameters.
The future work refers to the treatment centre of aged patients as a potential
beneficiary of such systems.
84 Machine Learning Algorithms and Applications in Engineering

Andrea et al. [42] designed a LSTM model for the video graphic detec-
tion of autism spectrum disorder. The study, done on 20 individuals with a
diagnosis of autism, fed a short video clip showing both healthy and autistic
participants to the neural network. The experiment consisted of 48 trials of
neutral hand gestures of grasping an object, which was filtered for discrep-
ancies. The evaluation for a thresholding value ranging from 0.5 to 0.95 gave
subsequent accuracy values of 83.33 percent and recall of 85 percent. The
frameworks have been deemed as progressive comparable to state-​of-​the-​art
LSTM systems along with the publication of the video data set for public use.
Ahmedt-​Aristizabal et al. [43] pursued computer vision to enumerate a
facial expression gesture analysis system for the early detection of epilepsy.
A system consisting of a video feed recording blinking, chewing, gazing
and motion of the jaw at 25 frames, a R-​CNN trained with the WIDER
data set and a Deep CNN for feature extraction followed by a LSTM with a
many-​to-​one layering for the evaluation. For result generation, a multi-​fold
cross-​validation method was applied, which measures both reliability and
consistency of the system. The results discussed show an average accuracy
of 96.58 percent with an AUC of 0.9926. This MTLE-​based detection has been
considered a success by the authors; however, the absence of large public
datasets of seizure facial expressions leaves room for overfitting-​based bias
in the results, as addressed in the paper itself.
Ye [44], in his paper, amalgamates the 3D conventional neural network
with the multi-​modal feature of the recurrent neural network to gain a
superior network. In their experiments, the C2D model is leveraged to
extract metric data from short video clips of ASL communication. Using
a greedy approach to stitch together the clips with the highest confidence
score, the proposed hybrid framework called 3DRCNN can train and work
with any of the multitude of ASL present in use. For the data set, the team
created a modified Kinect sensor in association with the ASL Association
and created short clips of 99 different hand gestures as well as 100 short
one-​minute video clips of sentences. After training on the Sports1M dataset,
testing showed an accuracy of 69.2 percent in case of person-​dependent
comparison and 65.2 percent in case of person-​independent analysis. Both
scores were compared and undefeated by two other systems, LRCN and
only C3D methods.
In another paper by Rastagoo [45], the team started by building a data set
by only 10 participants who provided 10,000 RGB videos of over 90 hand
gestures in the Persian language. For the model, a pipeline system starts from
a CNN to extract the heatmap of the hand using key points and boundary
boxes projected a skeleton of the hand on a plane that feeds it to the proposed
framework named 3DCNN. A total of five such parallel streams generates
a composite modal joint heatmap, which is processed via a LSTM network.
All possible combinations of the data streams with the LSTM, including
modifying the number of streams, has been recorded with the best result
A Comparative Analysis of Techniques in Gesture Detection 85

of 99.80 percent accuracy being crowned to the system of prefiltering with


multi-​view projection using five streams. Multiple comparisons with the
state-​of-​the-​art systems have been made, in which 2DCNN is taken into con-
sideration. In all the cases, the proposed framework scored competitively.
Gomez-​Donozo et al. [46] created a CNN-​based hand pose approximation
system with a mean error less than 5 mm. This system comprising of two
CNN is able to accurately predict the position of all joints of a user’s hand in
the spatial coordinate system from the viewpoint of any camera. The author
has further claimed that the system is comparable to a Leap Motion device.
The data set used for the training and testing of this system was a novel
dataset named LSMVHandPoses comprising of 21 sequences in four colours
for each of the 20,500 frames, which have been manually annotated and
compared with the predicted 2D computerised system. The paper mentions
the drawbacks of jitters being introduced in long video clips and suggests to
resolve this issue in the future iterations.
Another 3D hand model estimation pipeline was proposed by Spurr et al.
[47]. A cross-​modal training system composed of semi-​supervised learning-​
based statistical hand model representation in the spatial domain could be
used to recreate the position of the hand from the RGB channels of the video
feed. For the evaluation, no new data sets were created, and the publicly
available data sets Stereo Hand Pose Tracking Benchmark [48] (STB) and the
Rendered Hand Pose Dataset (RHD) [49] were used. The paper is too detailed
and the intricacies of the KL divergence framework complicated to under-
stand, but the results section is well adorned with over 30 images of detailing
the different inputs at different stages for 12 images.
Simon et al. [50] performed two major achievements in their works. First,
the main work focuses on creating a fine-​tuned full body key point detection
system using multiple cameras in tandem, and the corollary of generating an
automated key point marking system for automated reconstruction of a music
performer. This framework operates using the edge detection techniques of
multi-​view bootstrap followed by making of keyframes and training a deep
neural network. The only evaluation done was comparison of hand gestures,
and hand object detection using a Max3 data set against the salient point
model proposed in Ref. [51]. Although the framework shows up to 35 percent
improvements, the low range of numbers and lack of comparison against
other frameworks leaves many avenues unexplored. The system is a novel
approach to an already available technique that might be useful in detecting
full body gestures in nonverbal communications.
Haque [52], along with his research team, designated a discriminative
model for the prediction of the 3D model of a human from one full frame
still image. The primary goal of the study was to overcome the drawbacks of
the local patch-​based spatial arrangement techniques and ultimately achieve
viewpoint invariance. An iterative remedial approach was implemented
wherein the error feedback would lead to refinement in subsequent passes.
86 Machine Learning Algorithms and Applications in Engineering

To make the connections more direct with a linear relation to feedback of


the previous iterations, the framework used a LSTM as a filtering technique
to further reduce temporal dependency and access the tacit layers directly.
Evaluation was performed using the public Stanford EVAL data set, which
was compared against random forests [53], random tree walks (RTW) [54]
and iterative error feedback (IEF) [55] based on the percentage of correct key
points (PCKh) metric. The performance was slightly better than the compe-
tition and the highest increase in mean scores was found to be in the case
of the lower abdomen, while the head gave the least increase according to
the metrics. RF was not a suitable choice for contrasting the method as it
scored 40 percent lower than both the proposed framework and the other
two frameworks, which puts it out of the league of the other frameworks.
Another feature of the framework, ‘Glimps’ added more data with each iter-
ation along the pipeline and provided a noticeable increase in sharpness of
the heatmaps presented.
A simple system for feature extraction with gesture detection via a neural
network was presented by Jadooki [56]. The distinguishing factor of this
paper is that instead of stating all the systems used and the modalities
considered, they also provide insightful information about the importance
of the difference as well as the effect on results when the parameters are
changed. Simple RGB-​based image manipulating tools are used to extract the
features of the hand and the palm. The effects of lighting and the background
intensity has also been mentioned briefly. A simple neural network equipped
with a custom-​modified discrete cosine transform (DCT) method is used for
the feature processing and subsequent result generation. The effect of the
multiple layers and the layers that contribute the most to the accuracy of the
results have been provided along with graphs of the accuracy over iterations
for different tweaks to the variables to the DCT. The results conclude that
increasing certain mentioned layers cause linear reduction in the prediction
of the hand gesture, till 10,000 iterations of training, after which the frame-
work stagnates.
Priya et al. [57] has proposed the framework of a project to implement
an ASL-​based text input in laptop systems. This is a bit different from the
other studies already discussed, because instead of focussing on the intri-
cacies of the parameters and the pooling of layers of the network, the paper
mainly focuses on the requirements of the system to be able to work in real-​
world applications. Many techniques that are used for the extraction of
hand positions implemented deep inside the neural networks such as the
convex hull, erosion, black top hat and skeletonize have been studied and
the importance evaluated. Finally, the paper suggests methods to implement
the gesture as a virtual mouse input via the camera and gives a brief of the
theoretical aspects of the project.
The paper drafted by Smrita [58] outlines the use of a radio frequency-​
based microcontroller-​enabled home automation system that can understand
A Comparative Analysis of Techniques in Gesture Detection 87

gesture commands. The system uses a PIC microcontroller in tandem with


an accelerometer that is attached to the wrist as a wearable device to gather
the inputs. The microcontroller can converse with various relays using radio
frequency, hence there is no need to be in the same room as the appliance
that is to be controlled. Currently the system is rudimentary, with the system
being able to turn appliances on or off. With proper software support, this
system should be capable of carrying out more complex instruction like
setting reminders and alarms. The study aims that the system benefits para-
lytic patients with limited motor functionality, but it can also be applicable to
any case where only gesture is the primary means of communication.

5.5 Conclusion
Throughout this chapter, discussions have been made with the primary
focus on using gesture detection frameworks that have been proposed,
developed and implemented in real-​world scenarios. Gesture detection has
many uses, starting from the reasons mentioned in the first sections of the
chapter to newer intuitive implementation that have not been discovered
till date. This is not surprising considering the fact that in this day where
digital technology is widespread into every nook and cranny of the society,
the access to camera and the ability to use it is extremely trivial. As such,
anyone can leverage the fruits of the research done in the fields of gesture
analysis, with a main focus still remaining in the domain of ‘American sign
language’, which is one of the most widely used techniques to communicate
using palm and hand gestures.
In this chapter, the different challenges with the acquisition and processing
of data sets are discussed with certain examples detailing the reason for
it. Thereafter, the different requirements along the process of developing
usable and efficient frameworks from the system has also been perused over,
followed by the application of gesture analysis for the benefit of the society.
As previously stated, it is obvious that the most important and obvious use of
this technique is in developing tools for communicating with those who are
challenged in some form; however, this technique can also be modified for use
in novel ways, such as the detection and treatment of autism. Another similar
field that got a brief mention was the use in detection of sleep and fatigue
levels in students attending online classes as well as making 3D models of
complex actions from simple images. There is no doubt that with time, these
techniques would get refined by the work of the community, leading to other
fruitful discoveries.
Finally, an extended literature review of the current techniques that have
something unique to offer have been compiled in the review section. Instead
88 Machine Learning Algorithms and Applications in Engineering

of just stating the work done in those studies, relevant details such as the data
sets used, the deep learning techniques implemented for the development of
the pipelines, as well as the outcomes and comparative analysis performed
have been condensed and mentioned to aid in quick retrieval of important
information without access to the original work. A few minor observations
have been made: among the techniques evaluated, a large number of them
relied on LSTM to finalize and unify the data in the final layer of the models.
The use of deep learning-​based techniques has seen more prevalence, with
normal CNN being used mostly in extracting features from image and video
data sets. Most of the studies also had to rely on creating their own data sets,
which they have subsequently made public, which can only be seen as a huge
asset to the entire research community.

References
[1] Bellugi U, Fischer S (1972) A comparison of sign language and spoken language.
Cognition 1: 173–​200.
[2] Yang R, Sarkar S, Loeding B (2010) Handling movement epenthesis and hand
segmentation ambiguities in continuous sign language recognition using nested
dynamic programming. IEEE Trans Pattern Anal Mach Intell 32: 462–​477.
[3] Wu Y, Huang TS (1999) Vision-​ based gesture recognition: a review.
In: International Gesture Workshop, Springer, pp. 103–​115.
[4] Wu Y, Huang TS (1999) Human hand modeling, analysis and animation in the
context of HCI. In: Image Processing, ICIP 99. Proceedings. 1999 International
Conference, IEEE, pp. 6–​10.
[5] Starner T, Weaver J, Pentland A (1998) Real-​time American sign language recog-
nition using desk and wearable computer-​based video. IEEE Trans Pattern Anal
Mach Intell 20:1371–​1375.
[6] Grobel K, Assan M (1997) Isolated sign language recognition using hidden
Markov models. In: Systems, Man, and Cybernetics, 1997. Computational cyber-
netics and simulation. 1997 IEEE International Conference, IEEE, pp. 162–​167.
[7] Vogler C, Metaxas D (1998) ASL recognition based on a coupling between
HMMs and 3D motion analysis. In: Computer Vision, 1998. Sixth international
conference, IEEE, pp. 363–​369.
[8] Vogler C, Metaxas D (1997) Adapting hidden Markov models for ASL recogni-
tion by using three-​dimensional computer vision methods. In: Systems, Man, and
Cybernetics, Computational cybernetics and simulation. 1997 IEEE international
conference, IEEE, pp. 156–​161.
[9] Suarez J, Robin R (2012) Hand gesture recognition with depth images: A review.
IEEE RO-​MAN: The 21st IEEE International Symposium on Robot and Human
Interactive Communication Spet.
[10] Ferrone A, Maita F, Maiolo, and Arquilla M (2016) Wearable band for hand
gesture recognition based on strain sensors. IEEE RAS/​EMBS International
Conference on Biomedical Robotics and Biomechatronics 26–​29 June.
A Comparative Analysis of Techniques in Gesture Detection 89

[11] Plouffe G, Ana-​Maria C. Static and dynamic hand gesture recognition in depth
data using dynamic time warping, IEEE Transactions on Instrumentation and
Measurement, Vol. 65, no. 2, February 2016.
[12] Wei Lu, Zheng Tong, Jinghui Chu. Dynamic hand gesture recognition with leap
motion controller. IEEE signal Processing Letters, Vol 23, No. 9, September 2016.
[13] Alani AA, Georgina Cosma, Aboozar and McGinnity TM. Hand gesture recog-
nition using an adapted convolutional neural network with data augmentation.
4th IEEE International Conference on Information Management, 2018.
[14] Avola D, Bernardi M, Cinque L, Luca Foresti G, Massaroni C. Exploiting recur-
rent neural networks and leap motion controller for the recognition of sign lan-
guage and semaphoric hand gestures. Journal of Latex Class Files, Vol. 14, No 8,
August 2015.
[15] Zabulis X, Baltzakis H, Argyros A. Vision-​based hand gesture recognition for
human–​computer interaction. Gesture, 1–​56, 2009.
[16] Chakraborty B, Sarma D, Bhuyan M, Macdorman K. Review of constraints on
visionbased gesture recognition for human–​computer interaction (2018).
[17] Bauer, B., Karl-​Friedrich, K. (2002) Towards an automatic sign language recog-
nition system using subunits. In: Wachsmuth, I., Sowa, T. (eds.) GW 2001. LNCS
(LNAI), vol. 2298, pp. 64–​75. Springer, Heidelberg.
[18] Zhu, Y., Yang, Z., & Yuan, B. (2013, April). Vision based hand gesture recognition.
In 2013 International Conference on Service Sciences (ICSS) (pp. 260–​265). IEEE.
[19] Chakraborty, B. K., Sarma, D., Bhuyan, M. K., & MacDorman, K. F. (2018).
Review of constraints on vision-based gesture recognition for human–​computer
interaction. IET Computer Vision, 12(1), 3–​15.
[20] Liang, RH, Ouhyoung M. (1998, April). A real-​time continuous gesture recogni-
tion system for sign language. In Proceedings third IEEE international confer-
ence on automatic face and gesture recognition (pp. 558–​567). IEEE.
[21] Morency, LP, Quattoni A, Darrell T. (2007, June). Latent-​dynamic discriminative
models for continuous gesture recognition. In 2007 IEEE conference on com-
puter vision and pattern recognition (pp. 1–​8). IEEE.
[22] Morency LP, Christoudias CM, Darrell, T. (2006, November). Recognizing gaze
aversion gestures in embodied conversational discourse. In Proceedings of the
8th international conference on Multimodal interfaces (pp. 287–​294).
[23] Krauss, RM, Chen Y, Chawla P. (1996). Nonverbal behavior and nonverbal
communication: What do conversational hand gestures tell us?. In Advances in
Experimental Social Psychology (Vol. 28, pp. 389–​450). Academic Press, London.
[24] Vega K, Cunha M, Fuks H. (2015, March). Hairware: the conscious use of
unconscious auto-​contact behaviors. In Proceedings of the 20th International
Conference on Intelligent User Interfaces, pp. 78–​86.
[25] Mitra S, Acharya T. Gesture recognition: a survey. IEEE Trans. Syst. Man Cybern.
Part C Appl. Rev. 37(3), 311–​324 (2007).
[26] Wobbrock JO, Morris MR, and Wilson AD. (2009). User-​defined gestures for sur-
face computing. Proc. CHI ‘09, 1083–​1092.
[27] Morris MR, Wobbrock JO, Wilson AD. (2010). Understanding users’ preferences
for surface gestures. Proc. GI ‘10, 261–​268.
[28] Avrahami D, Hudson SE, Moran TP, Williams BD. (2001). Guided gesture
support in the paper PDA. Proc. UIST ‘01, 197–​198.
[29] Bigdelou A, Schwarz L, Navab. (2012). An adaptive solution for intra-​operative
gesture-​based human-​machine interaction. Proc. IUI ‘12, 75–​83.
90 Machine Learning Algorithms and Applications in Engineering

[30] Döring T, Kern D, Marshall P, Pfeiffer M, Schöning J, Gruhn V, Schmidt, A. (2011).


Gestural interaction on the steering wheel –​reducing the visual demand. Proc.
CHI ’11, 483–​492.
[31] Fiebrink R, Cook PR, Trueman D. (2011). Human model evaluation in interactive
supervised learning. Proc, CHI ’11, 147–​156.
[32] Yang S, Liu J, Lu S, Er MH, Kot AC. (2020, August). Collaborative learning of
gesture recognition and 3D hand pose estimation with multi-​order feature ana-
lysis. In Proceedings of the European Conference on Computer Vision (ECCV).
[33] Clarke JS, Llewellyn N, Cornelissen J, Viney R. (2019). Gesture analysis and
organizational research: the development and application of a protocol for nat-
uralistic settings. Organizational Research Methods, 1094428119877450.
[34] Ramos-​Cabo S, Vulchanov V, Vulchanova, M. (2019). Gesture and language tra-
jectories in early development: An overview from the autism spectrum disorder
perspective. Frontiers in Psychology, 10, 1211.
[35] Ginosar S, Bar A, Kohavi G, Chan C, Owens A, Malik J. (2019). Learning indi-
vidual styles of conversational gesture. In Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, pp. 3497–​3506.
[36] Li W, Jiang F, Shen R. (2019, May). Sleep gesture detection in classroom monitor
system. In ICASSP 2019–​ 2019 IEEE International Conference on Acoustics,
Speech and Signal Processing (ICASSP), pp. 7640–​7644. IEEE.
[37] Trujillo JP, Vaitonyte J, Simanova I, Özyürek A. (2019). Toward the markerless
and automatic analysis of kinematic features: A toolkit for gesture and movement
research. Behavior Research Methods 51(2): 769–​777.
[38] Rastgoo R, Kiani K, Escalera, S. (2020). Sign language recognition: A deep survey.
Expert Systems with Applications, 113794.
[39] Mustafa M. (2020). A study on Arabic sign language recognition for differently
abled using advanced machine learning classifiers. Journal of Ambient Intelligence
and Humanized Computing, 1–​15.
[40] Taskiran M, Killioglu M, Kahraman N. (2018, July). A real-​time system for
recognition of American sign language by using deep learning. In 2018 41st
International Conference on Telecommunications and Signal Processing (TSP)
(pp. 1–​5). IEEE.
[41] Heydarian H, Rouast PV, Adam MT, Burrows T, Collins CE, Rollo ME (2020).
Deep learning for intake gesture detection from wrist-​worn inertial sensors: The
effects of data preprocessing, sensor modalities, and sensor positions. IEEE
Access, 8, 164936–​164949.
[42] Zunino A, Morerio P, Cavallo A, Ansuini C, Podda J, Battaglia F, ... Murino
V. (2018, August). Video gesture analysis for autism spectrum disorder detec-
tion. In 2018 24th International Conference on Pattern Recognition (ICPR)
(pp. 3421–​3426). IEEE.
[43] Ahmedt-​Aristizabal D, Fookes C, Nguyen K, Denman S, Sridharan S, Dionisio,
S. (2018). Deep facial analysis: A new phase I epilepsy evaluation using com-
puter vision. Epilepsy & Behavior, 82, 17–​24.
[44] Ye Y, Tian Y, Huenerfauth M, Liu J. (2018). Recognizing American sign language
gestures from within continuous videos. In Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition Workshops (pp. 2064–​2073).
[45] Rastgoo R, Kiani K, Escalera S. (2020). Hand sign language recognition using
multi-​view hand skeleton. Expert Systems with Applications, 113336.
A Comparative Analysis of Techniques in Gesture Detection 91

[46] Gomez-​Donoso F, Orts-​Escolano S, Cazorla M. (2019). Accurate and efficient


3D hand pose regression for robot hand teleoperation using a monocular RGB
camera. Expert Systems with Applications, 136, 327–​337.
[47] Spurr A, Song J, Park S, Hilliges O. (2018). Cross-​modal deep variational hand
pose estimation. In Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition (pp. 89–​98).
[48] Zhang J, Jiao J, Chen M, Qu L, Xu X, Yang Q. (2017, September). A hand pose
tracking benchmark from stereo matching. In 2017 IEEE International Conference
on Image Processing (ICIP) (pp. 982–​986). IEEE.
[49] Zimmermann C, Ceylan D, Yang J, Russell B, Argus M, Brox T. (2019).
Freihand: A dataset for markerless capture of hand pose and shape from single
RGB images. In Proceedings of the IEEE International Conference on Computer
Vision (pp. 813–​822).
[50] Simon T, Joo H, Matthews I, Sheikh Y. (2017). Hand keypoint detection in single
images using multiview bootstrapping. In Proceedings of the IEEE conference
on Computer Vision and Pattern Recognition (pp. 1145–​1153).
[51] Tzionas D, Ballan L, Srikantha A, Aponte P, Pollefeys M, Gall J. Capturing hands
in action using discriminative salient points and physics simulation. IJCV, 2016.
[52] Haque A, Peng B, Luo Z, Alahi A, Yeung S, Fei-​Fei, L (2016, October). Towards
viewpoint invariant 3d human pose estimation. In European Conference on
Computer Vision (pp. 160–​177). Springer, Cham.
[53] Shotton J, Fitzgibbon A, Cook M, Sharp T, Finocchio M, Moore R, ... Blake
A. (2011, June). Real-​time human pose recognition in parts from single depth
images. In CVPR 2011 (pp. 1297–​1304). IEEE.
[54] Yub Jung H, Lee S, Seok Heo Y, Dong Yun I. (2015). Random tree walk toward
instantaneous 3d human pose estimation. In Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition (pp. 2467–​2474).
[55] Carreira J, Agrawal P, Fragkiadaki K, Malik J. (2016). Human pose estimation
with iterative error feedback. In Proceedings of the IEEE conference on com-
puter vision and pattern recognition (pp. 4733–​4742).
[56] Jadooki S, Mohamad D, Saba T, Almazyad AS, Rehman A. (2017). Fused features
mining for depth-​based hand gesture recognition to classify blind human com-
munication. Neural Computing and Applications, 28(11): 3285–​3294.
[57] Priya, T., Chanukah, P., & Vanusha, D. (2019). Hand gesture to audio based com-
munication system for blind people.
[58] Rumana, T. A. (2015). HAND GESTURE BASED HOME AUTOMATION FOR
VISUALLY CHALLENGED. HAND, 2(4).
6
Brain–​Computer Interface for Dream
Visualization using Deep Learning

Brijesh K. Soni1 and Akhilesh A. Waoo2


1
Assistant Professor, Department of Computer Science and Technology,
AKS University, Satna, MP, India
2
Head and Associate Professor, Department of Computer Science and Technology,
AKS University, Satna, MP, India

CONTENTS
6.1 Dream: Introduction..................................................................................... 94
6.1.1 Theories of Dream............................................................................ 94
6.1.1.1 Wish-Fulfillment Theory................................................... 95
6.1.1.2 Information-Processing Theory....................................... 95
6.1.1.3 Activation Synthesis Theory............................................ 95
6.1.1.4 Physiological Function Theory........................................ 96
6.1.2 Origin of Dream................................................................................ 96
6.1.2.1 Dream versus Sleep........................................................... 96
6.1.2.2 Dream versus Memory...................................................... 97
6.1.2.3 Human Visual Cortex........................................................ 98
6.1.2.4 Hippocampus and Amygdala.......................................... 99
6.2 Brain–Computer Interface......................................................................... 100
6.2.1 BCI Tools.......................................................................................... 101
6.2.1.1 Electroencephalography (EEG)...................................... 101
6.2.1.2 Electrocardiography (ECG)............................................ 101
6.2.1.3 Magnetic Resonance Imaging (MRI)............................. 102
6.2.1.4 Magnetoencephalography (MEG)................................. 102
6.2.2 Procedure......................................................................................... 103
6.2.2.1 Signal Acquisition............................................................ 103
6.2.2.2 Pattern Recognition......................................................... 103
6.2.2.3 Pattern Classification....................................................... 104
6.2.2.4 Command Generation..................................................... 104

DOI: 10.1201/9781003104858-6 93
newgenprepdf

94 Machine Learning Algorithms and Applications in Engineering

6.3 Deep Learning............................................................................................. 104


6.3.1 Deep Learning Models.................................................................. 105
6.3.1.1 Deep-Belief Network (DBN)........................................... 105
6.3.1.2 Recurrent Neural Network (RNN)................................ 106
6.3.1.3 Convolutional Neural Networks (CNN)...................... 106
6.3.1.4 Gradient-Adversarial Network (GAN)........................ 106
6.3.2 Deep Learning Frameworks......................................................... 107
6.3.2.1 Keras.................................................................................. 107
6.3.2.2 Caffe................................................................................... 108
6.3.2.3 PyTorch.............................................................................. 108
6.3.2.4 MATLAB........................................................................... 109
6.4 Conclusion....................................................................................................110

6.1 Dream: Introduction
Philosophers around the world believe that dream is one of the most curious
things in human beings [1]. The dream is the succession of images and
sounds occurring inside the mind during sleep [2]. Researchers have found
that visual and auditory stimuli are active during certain stages of sleep
within the brain regions [3]. Some researchers found in their research that
the hippocampus, amygdala, visual cortex, auditory cortex, and motor cortex
are the major regions that are active during the dreaming of a sleeping object
[4]. However, most of the oneirologists committed in various articles that
dream originates from the centrally located brain stem area but associated
with other important cortical areas also.
There are various futuristic research ideas in my mind related to dreams,
and belief says that if it will be possible to implement some of them, it might
be beneficial for our community. Suppose a dreamer may spend six to eight
hours every night sleeping, he is also spending almost an equal time in the
workplace during daytime for his survival [5]. So, the question is, is it pos-
sible to make use of that time spent sleeping in any way. The dream might
be a solution for making sleeping time useful; it means if it is possible to con-
trol dream activities and contents, it will help humans to grow financially by
increasing our effective work hours.

6.1.1 Theories of Dream
Since the early sixteenth century, scientists are striving to know the fun-
damental causes of why dream occurs, but still, no one sure about it [6].
There are various philosophical theories proposed by various scientists and
philosophers from time to time, and some popular theories of the dream
(Figure 6.1) are briefly described here.
Brain–Computer Interface for Dream Visualization 95

Dream-Theories

Whish Fulfillment Information Activation Physiological


Theory Processing Theory Syntheis Theory Theory

FIGURE 6.1
Various theories of dreaming.

6.1.1.1 Wish-​Fulfillment Theory
Wish-​fulfillment theory was proposed by an Austrian oneirologist, Sigmund
Freud. He believed that dream contents are related to the wishes or desires of
the dreamer. Generally, a dreamer sees contents into a dream, which are his
incomplete or unfilled desires in his real life [7]. The most common example
of such a type of dream is sexual arousal during sleep when an adult dreamer
might be involved in sexual or romantic activities in his dream. Some other
examples are feeling sadness and happiness, participating in any traditional
ceremony, visiting new places or any ancient place and finding voluble things
or properties [8].

6.1.1.2 Information-​Processing Theory
This theory deals with the process of computation like a computer system to
get input from the external environment and process the data and finally gen-
erate output. This theory comprises its data transmission path and processes
in the human brain just like computer hardware circuits and physical devices
are organized in a computer architecture [9].
It is well known that data and information inside the computer system
travel through busses and from one device to another device. In the same
manner, neurotransmitters travel through the axons and dendrites of neuron
cells in the nervous system or biological neural network [10]. Further, the
central part of neuron cells, which is also known as soma, is responsible for
collecting and modifying information in the form of chemical composition or
neurotransmitters. The same concept is also applicable in the case of dream
content generation inside the brain [11].

6.1.1.3 Activation Synthesis Theory


This theory is developed based on identifying active brain regions during
dreaming. Neuroscientists have seen in experiments that some spe-
cific regions are active and produce various types of wave signals, which
support various features in dream contents [12]. In-​depth, it can be seen that
96 Machine Learning Algorithms and Applications in Engineering

in different regions at different states of mind, various types of signals are


generated [13].

6.1.1.4 Physiological Function Theory


This theory of dreaming deals with periodic stimulation of the nervous
system. Brain stimulates periodically during the dreaming and strengthens
the information stored inside the brain as memory, whatever the dreamer
has seen during the waking period [14]. Generally, it happens at every stage
of the rapid eye movement (REM) sleep every night. That means the brain
works intuitively for strengthening the memory and data pattern inside the
brain [15].

6.1.2 Origin of Dream
Scientists undertook imaging of the many regions of the brain that are active
while dreaming, starting from the brain stem to the cortex through the
thalamus [16]. Most of the observations show that brain regions are active
similar to waking stages. Suppose you are seeing any visual imagery during
dreaming, it relates to the visual cortex in the occipital lobe of the brain. In
the same way, the other cortex is also related to concerned activities [17].
The cortical regions hippocampus and amygdala are responsible for feeling
emotions and sensation activities [18].

6.1.2.1 Dream versus Sleep


Sleep is the unconscious state of mind. Sleeping and awakening is a cyclic
process in our daily life. Most of the physiologists believe that our whole
body, including the brain, needs relaxation, which is possible due to some
chemical reactions. That’s why every animal takes rest by way of sleeping
[19]. Generally, a human being sleeps six to eight hours every night. Further,
neuroscientists found that at some stages of dreaming, human brain becomes
more active instead of going into a relaxation mode [20]. Various activities
are happening during sleeping, but the dream is the most effective activity in
our sleeping brain.
Before going into depth about dreaming, a little more about the basics of
sleep needs to be discussed. Pieces of literature recognize that sleep can be
classified into two broad categories, known as REM sleep and NREM sleep
[21]. The REM sleep stage is wholly responsible for dreams occurring in
the brain. A dreamer’s brain is more active in the REM sleep stage due to
dreaming activities (Figure 6.2). Table 6.1 shows time distribution for various
sleep stages.
Brain–Computer Interface for Dream Visualization 97

FIGURE 6.2
Neocortical activations during REM sleep.
(Source: www.menta​lhea​lths​cien​ces.com/​)

TABLE 6.1
Time Distribution for Various Sleep Stages

Stages Episode-​1 Episode-​2 Episode-​3 Episode-​4 Episode-​5 Total

NREM
Stage-​1  5  5  5  5  5  25
Stage-​2 50 40 40 30 20 180
Stage-​3 20 20 10 10 10  70
Stage-​4  5  5  5  5  5  25
REM 10 20 30 40 50 150
Total 90 90 90 90 90 90×5=​450+​25*=​480

6.1.2.2 Dream versus Memory


Researchers have found that that dreaming is a reasonable activity of memory
formation. As we know very well that our brain stores information inside
the brain in the form of a chemical pattern, it is further known as memory
[22]. Memory is classified into two categories as temporary memory and per-
manent memory. During dreaming, temporary memory might transform
into permanent memory. This process of memory transformation is referred
98 Machine Learning Algorithms and Applications in Engineering

FIGURE 6.3
Long-​term memory formation.
(Source: www.nat​ure.com)

to as memory consolidation or memory synthesis. However, the activation-​


synthesis theory of dreaming shows that dreaming is also responsible for
strengthening memory by analyzing and improving chemical compositions
[23]. Initially, temporary memory is formed in the hippocampus, but during
consolidation, it might be encoded and stored in the cortical regions of the
brain, which might be retrieved in the future. The hippocampus transfers
temporary memories to the associated neocortex for long-​ term storage
(Figure 6.3).
Now the importing concept about memory synthesis, which might be a
part of memory consolidation while dreaming, is memory encoding. Here
the process of memory encoding is discussed with the help of proper con-
struct and pattern. There are four types of memory encoding, discussed in
various articles and literature: visual, acoustic, elaborative, and semantic
[24]. Stimuli involved in the process of memory encoding may be the dream
content a dreamer can see frequently in his real-​time dreaming.

6.1.2.3 Human Visual Cortex


The visual cortex is a very crucial part of the human brain located in the
occipital lobe. All the visual stimuli are mapped inside the visual cortex.
Anatomically, visual cortex is divided into two parts and is located in the
left hemisphere and right hemisphere. Here the intentional focus is on
the visual cortex because our primary goal is to recognize and visualize
visual stimuli only occurring in a dream [25]. Most research engineers and
scholars also committed that whatever a dreamer sees in a dream might
be mapped within the visual cortex. Even the experimental psychology
research outcomes show that the visual cortex is fully active while dreaming
(Figure 6.4). However, human curiosity is to know that, if someone got
damaged in his visual cortex, then can he see any visual content in his
dream? [26].
Brain–Computer Interface for Dream Visualization 99

FIGURE 6.4
Visual cortex activation during REM sleep.
(Source: www.nat​ure.com)

FIGURE 6.5
Position of hippocampus inside the limbic system.
(Source: https://​med​imag​ery.com/​)

6.1.2.4 Hippocampus and Amygdala


Hippocampus is a dominating organ of the human brain where short-​term
or temporary memory is formed. It regulates motivation and learning,
but it plays a crucial role in memory encoding and consolidation also [27].
Physically, it is located in the central area of the brain (Figure 6.5), which
means it is a part of the limbic system and thalamus in the cerebral cortex,
and strongly connected with all other cortical regions for accessing and
forwarding information [28]. Researchers found that the hippocampus plays
a vital role in dreaming while a dreamer sees some past experiences and
memories in his dream.
Like the hippocampus, amygdala appears in two pieces in each part of the
hemisphere (Figure 6.6). It has a basic function in emotional memory [29].
100 Machine Learning Algorithms and Applications in Engineering

FIGURE 6.6
Position of amygdala inside the limbic system.
(Source: www.reba​lanc​ecli​nic.co.nz)

Researchers believe that like the hippocampus and other cortical regions,
amygdala is also active during dreaming and participates in emotional activ-
ities in dreaming. A dreamer can feel fear, anger, anxiety, and violence due to
the presence of amygdala [30].

6.2 Brain–​Computer Interface
At the early stage of this research work, core concepts and anatomical
classifications on brain–​computer interface (BCI) were published. The pre-
sent chapter focused on an important application in the domain of dream
physiology (oneirology) [31]. The previous section described important brain
organs and activities associated with dreaming, and this section discusses
some important modalities and techniques frequently used in the domain of
brain–​computer interface [32].
In clinical bionics, the brain–​computer interface is a very popular term
among biomedical engineers and researchers. In general, the brain–​computer
interface is a communication pathway between the human brain and the com-
puter system [33]. Fundamentally, this domain has two broad subdomains
where a researcher must focus on modalities and procedures.

• Modalities mean those hardware components which are used to extract


signals from the brain invasively or noninvasively.
• The procedure means the data analytics process using software
components or modules [34].
Brain–Computer Interface for Dream Visualization 101

Further, these hardware modalities and procedural implementation of soft-


ware modules are integrally known as brain–​computer interface technology.
This technology is a very important part of this chapter where it requires
to extract dream contents from the human brain and visualize them. First,
modalities that can be used may be based on electrical activities or meta-
bolic activity measurements from the brain, and it may be in an individual
or hybrid form also [35]. Second, detailed procedure feature extraction and
command generation for handling or performing any specific activity is
implemented in the form of software modules. Now the point is that these
software modules are based on deep learning concepts [36]. However, deep
learning theory and respective models with important frameworks are
described in the last section of this chapter.

6.2.1 BCI Tools
Researchers are using a tool or modality for extracting signals/​images from
the brain. Here, some popular modalities are briefly described.

6.2.1.1 Electroencephalography (EEG)
Electrical activity measurement is the most popular study among brain
researchers. Electroencephalography (EEG) is one of the major technologies
for electrical signal acquisition from the human brain by placing electrodes
noninvasively on the scalp during the experiment [37]. EEG has various
applications in the domain of brain–​computer interface such as medical diag-
nosis, game playing, skill improvement, and psycho-​physiological analysis.
This information can be extracted from various neocortical regions of the
brain. As stated in the first section, dreaming is directly associated with
memory consolidation happening inside the various neocortical regions of
the brain [38]. EEG modality is used for tracking electrical activities occurring
during memory consolidation during REM sleep. Electrodes must be placed
systematically (10–​20 system) in cortical regions. Some major cortical regions
are the visual cortex, auditory cortex, motor cortex, and also thalamus region
[39]. That means we have to measure electrical activities from the frontal
lobe, occipital lobe, and temporal lobe for getting different features of dream
contents.

6.2.1.2 Electrocardiography (ECG)
Electrocardiography is also used to measure electrical activities inside the
human body, but in general, it is used for recording and analyzing electrical
activities of the heart and other associated organs. However, it cannot be used
directly for brain diagnosis, but sometimes researchers try to study brain-​
instructed activities occurring in other associated organs of the body [40].
102 Machine Learning Algorithms and Applications in Engineering

Psychologists and dream researchers believe that during dreaming, some


other organs of the body are also active and may perform some activities
other than brain regions. Likely some researchers are actively engaged in
dream research using ECG modalities instead of EEG [41]. Nightmares are
occur in this type of a dreamer who actively shows body movement during
dreaming, and this is directly related to motor activities instructed by the
motor cortex [42].

6.2.1.3 Magnetic Resonance Imaging (MRI)


Researchers committed that this is a modern technology for brain imaging
compared to EEG/​ECG. In this technology, the concept of magnetism is
applied instead of electrical activities [43]. Magnetic resonance imaging
(MRI) is a radiology-​based scanning technique that uses the concept of mag-
netism, radio waves, and computation to produce images of body structures.
This is a tube-​like structure surrounded by a large circular magnetic compo-
nent [44]. The dreamer is placed on a movable bed that is inserted into the
magnetic cylinder. The associated magnet creates a strong magnetic field that
aligns the protons of hydrogen atoms, which are then exposed to a beam
of radio waves [45]. This spins the various regions of the brain, and they
produce a faint signal that is detected by the receiver portion of the scanner.
A software tool processes the receiver information, which produces an image.
The generated image and resolution are quite detailed, and it can detect tiny
metabolic changes of structures within the cortical regions [46].
Dream researchers use MRI for detecting active regions within the brain
during dreaming. Using the MRI technique, very small active regions can
also be detected easily and precisely. Metabolic conditions within the neuron
cells show the activities happening during dreaming in REM sleep [47].

6.2.1.4 Magnetoencephalography (MEG)
The basic concept of MEG is also magnetism, but unlike MRI, in this tech-
nology any external magnet is not used. It is based on the magnetic property
of blood. Magnetoencephalography, or MEG scan, is an imaging technique
that identifies brain activity and measures small magnetic fields produced in
the brain [48]. The scan is used to generate a magnetic image to pinpoint the
source of seizures. In this imaging technique, magnetic fields are detected by
extremely sensitive devices [49]. This device is known as superconducting
quantum interference device (SQID), frequently used to detect the mag-
netic field. The MEG scanner is used to detect and amplify magnetic signals
produced by the neurons of various active regions of the brain, and unlike
MRI, it doesn’t emit radiation or magnetic fields [50].
Comparative study of EEG, ECG, MRI and MEG signals acquisition modal-
ities, is shown in Table 6.2.
Brain–Computer Interface for Dream Visualization 103

TABLE 6.2
Comparative Study of Various Signals Acquisition Modalities

Parameter EEG ECG MRI MEG

Electrical Electrical Structural Magnetic


Reading Activity Activity Imaging Activity

Purpose Brain diagnosis Heart diagnosis Body diagnosis Brain diagnosis


Channels Multichannel Single-​channel/​ Multichannel Multichannel
multi-​channel
Cost ~2000–​3000 USD ~500-​1000 USD ~1000000 USD 5000-​10000 USD

6.2.2 Procedure
The brain–​ computer interface is a mediator method and procedure for
establishing communication between the dreamers. A BCI communication
system has to be built by following various activities and tools such as soft-
ware and hardware. In this section, a detailed description of developing a
BCI system with complete steps is presented.

6.2.2.1 Signal Acquisition
In the formal working procedure of a brain–​ computer interface system,
signal acquisition is the first phase where a hardware modality is used to
extract signals from the brain. Generally, hardware modality is associated
with the brain directly or indirectly [51]. The way of connection depends on
the type of modality; it may be wired or wireless and it may also be inva-
sive, semi-​invasive, and noninvasive type of connection. Now consider the
example of an EEG machine. So, it may be a wired or wireless system [52].
In early times, researchers used wired EEG systems, but nowadays, most
of the dream researchers also used mobile or wireless EEG modality. In the
case of EEG-​based experiment, an electrode cap having motile electrodes are
placed on the scalp. These electrodes read the electrical stimulus from the
brain regions, which depends on the number of channels designed in the
machine [53].

6.2.2.2 Pattern Recognition
Recorded raw data might have some noise and redundancy, so in this phase,
we have to identify patterns or features from the brain signals that have been
recorded. Pattern recognition is the task of analyzing the signals to distin-
guish significant signal features from general raw materials and representing
them in a standard form suitable for translation into commands during fea-
ture translation [54]. When the input signals to an algorithm are too large
and it is suspected to be redundant, then it can be transformed into a reduced
104 Machine Learning Algorithms and Applications in Engineering

set of features. Sometimes, unnecessary portions of the raw data set must be
removed, and the process of removing such type of data portion is known
as dimensionality reduction. The extracted features are expected to contain
valuable and relevant information so that the desired task can be performed
by using this reduced pattern instead of the complete source data [55].

6.2.2.3 Pattern Classification
It is the process of categorizing data based on their features to perform the
variability reduction of feature values. This stage classifies the extracted fea-
ture signals having different features into account [56]. The responsibility of
the feature classifier algorithm is to use the feature vector provided by the fea-
ture extractor to assign the object to a category of the feature [57]. However,
complete classification is often impossible, so a more common task is to
determine the probability for each of the possible categories of features. The
problem of the classification depends on the variation in the feature values
for certain objects in the same category relative to the variation between fea-
ture values for certain objects in different categories [58]. The variation of
feature values for certain objects in the same category may be due to the com-
plexity of features and may be due to noise in signals.

6.2.2.4 Command Generation
Finally, based on data evaluation, decisions could be taken by computer
for performing any specific task by generating appropriate command. This
stage performs command generation [59]. This stage translates the signals
into meaningful commands for any connected device. The classified feature
signals are translated by the feature translation algorithm, which converts
the feature signals into the appropriate commands for the specific operations
performed by the connected device [60]. In this context source feature,
signals are known as the independent variable and targeted device control
commands are known as the dependent variable. In the translation process,
the independent variable is converted into the dependent variable. Feature
translation algorithms may be linear or nonlinear by using statistical analysis
and neural networks, respectively [61].

6.3 Deep Learning
Intelligence is a key factor in the human brain; that’s why it is known as a
super creature of nature. It is well known that intelligence couldn’t occur
randomly within the muscular tissues, but it is only possible through the
Brain–Computer Interface for Dream Visualization 105

continuous process of learning. How a human brain solves problems of rec-


ognizing objects and associated patterns even in the case of complex images
or high-​volume data items [62]. In the same sense, simple algorithms are not
capable of solving such types of critical problems so that we need algorithms
having the highest capabilities for going into the deepest level and solving
such critical problems.
Dreamers around the world believe that mostly dream contents are not
easy to interpret and understand clearly. Deep learning provides efficient
and successful computational models based on the phenomenon of the
human nervous system [63]. These neural network models may be capable
of extracting and visualizing dream contents to study comparatively these
models with their efficiency and performance [64].
This section provides detailed discussions on various deep learning
models and possible frameworks for implementing these models and gener-
ating desired results. There are four models in this domain that provide com-
petitive performance with various data sets by adjusting their weights and
functions [65]. RNN and CNN are two popular models nowadays among the
data scientists for processing and evaluating data sets and generating appro-
priate outcomes. Gradient-​adversarial network (GAN) is another variant of
CNN having a hybrid computational capability [66].

6.3.1 Deep Learning Models


Deep learning models are deep neural networks that can be used for pro-
cessing textual and visual data sets. This section discusses the develop-
ment of various models for processing and analyzing real-​time brain signals
during REM sleep. These data might have various features such as auditory
data and visual data. Necessary deep neural network models are described
in this section.

6.3.1.1 Deep-​Belief Network (DBN)


This is the basic structure used for deep learning developed in the early
time, which was not more efficient, and now it is replaced by some new
and modern models. Deep belief networks are algorithms that use probabil-
ities and unsupervised learning to produce outputs. They are composed of
binary latent variables, and they contain both undirected layers and directed
layers [67].
Unlike other models, each layer in deep belief networks learns the entire
input. In convolutional neural networks, the first layers only filter inputs for
basic features, such as edges, and the later layers recombine all the simple
patterns found by the previous layers. Deep belief networks, on the other
hand, work globally and regulate each layer in order [68].
106 Machine Learning Algorithms and Applications in Engineering

6.3.1.2 Recurrent Neural Network (RNN)


This is the most popular model for natural language processing. It uses the
concept of recurrence using the backpropagation technique for adjusting
weights. In this type of network, the output from the previous step is fed
as input to the current step [69]. In traditional neural networks, all the
inputs and outputs are independent of each other, but in cases like when
it is required to predict the next word of a sentence, the previous words are
required and hence there is a need to remember the previous words [70].
Thus, RNN came into existence, which solved this issue with the help of a
hidden layer. The main and most important feature of RNN is the hidden
state, which remembers some information about a sequence [71].
RNN has a memory that remembers all information about what has been
calculated. It uses the same parameters for each input as it performs the same
task on all the inputs or hidden layers to produce the output. This reduces the
complexity of parameters, unlike other neural networks [72].

6.3.1.3 Convolutional Neural Networks (CNN)


This is the topmost neural network model nowadays among researchers and
industrialists. This is developed for image and video processing. It can take
in an input image, assign learnable weights, and biases to various aspects/​
objects in the image and be able to differentiate one from the other [73]. The
preprocessing required on CNN is much lower as compared to other classi-
fication algorithms. While in primitive methods filters are hand-​engineered,
with enough training, CNN can learn these filters/​characteristics [74].
The architecture of a CNN is analogous to that of the connectivity pattern
of neurons in the human brain and was inspired by the organization of the
visual cortex. Individual neurons respond to stimuli only in a restricted
region of the visual field known as the receptive field. A collection of such
fields overlaps to cover the entire visual area [75].

6.3.1.4 Gradient-​Adversarial Network (GAN)


This is the hybrid model developed by combining two models as a convo-
lutional model and a deconvolutional model. In GANs, there is a generator
and a discriminator. The generator generates fake samples of data and tries to
fool the discriminator [76]. The discriminator, on the other hand, tries to dis-
tinguish between the real and fake samples. The generator and the discrim-
inator are both neural networks and they both run in competition with each
other in the training phase [77]. The steps are repeated several times and in
this, the generator and discriminator get better and better in their respective
jobs after each repetition.
Comparative Study of DBN, RNN, CNN, and GAN Deep Learning Models
are shown in Table 6.3.
Brain–Computer Interface for Dream Visualization 107

TABLE 6.3
Comparative Study of Various Deep Learning Models

Parameter DBN RNN CNN GAN

Multilayer Multilayer
Type Staked RBM Perceptron Perceptron Hybrid Network

Application Image String processing Image Image processing


recognition processing
Vanishing Solved by chain Suffering from Solved by Relu Solved by
gradient rule vanishing Wasserstein
gradient loss
Dimensionality Possible Possible Possible Possible
reduction
Algorithm Unsupervised Unsupervised Supervised Unsupervised

6.3.2 Deep Learning Frameworks


Deep learning is an important element of medical imaging, which includes
statistics and predictive modeling. It is extremely beneficial to data scientists
who are tasked with collecting, analyzing, and interpreting large amounts of
data; deep learning makes this process faster and easier. At its simplest, deep
learning can be thought of as a way to automate predictive analytics. While
traditional machine learning algorithms are linear, deep learning algorithms
are stacked in a hierarchy of increasing complexity and abstraction [78]. Each
algorithm in the hierarchy applies a nonlinear transformation to its input and
uses what it learns to create a statistical model as output. Iterations continue
until the output has reached an acceptable level of accuracy. The number of
processing layers through which data must pass is what inspired the label
deep [79].

6.3.2.1 Keras
Google Brain developed TensorFlow, which is used as a backend for Keras.
Keras is an open-​source framework developed by Google engineer Francois
Chollet and it is a deep learning framework easy to use and evaluate our
models, by just writing a few lines of code [80]. Keras is the best framework
to start for beginners, and it was created to be user-​friendly and easy to work
with Python and it has many pre-​trained models—​VGG, Inception, among
others. Not only ease of learning, but in the backend, it supports Tensorflow
and is used in deploying models [81]. Keras was created to be user-​friendly,
modular, easy to extend, and to work with Python. The API was designed
for human beings, not machines, and follows best practices for reducing cog-
nitive load. Neural layers, cost functions, optimizers, initialization schemes,
activation functions, and regularization schemes are all standalone modules
that can be combined to create new models [82]. New modules are simple to
108 Machine Learning Algorithms and Applications in Engineering

add as new classes and functions. Models are defined in Python code, not
separate model configuration files.
The biggest reasons to use Keras stem from its guiding principles, primarily
the one about being user-​friendly. Beyond the ease of learning and ease of
model building, Keras offers the advantages of broad adoption, support for a
wide range of production deployment options, integration with at least five
backend engines TensorFlow, CNTK, Theano, MXNet, and PlaidML, and
strong support for multiple GPUs and distributed training. Further, Keras is
backed by Google, Microsoft, Amazon, Apple, Nvidia, Uber, and others [83].

6.3.2.2 Caffe
Convolutional architecture for fast feature embedding (Caffe) is the open-​
source deep learning framework developed by the University of Berkeley
AI Research Group. The framework is available as free open-​source software
under a BSD license. This framework supports both researchers and indus-
trial applications in Artificial Intelligence [84]. Caffe is a deep learning frame-
work characterized by its speed, scalability, and modularity. Caffe works
with CPUs and GPUs and is scalable across multiple processors. Most of
the developers use Caffe for its speed, and it can process 60 million images
per day with a single NVIDIA K40 GPU [85]. Caffe has many contributors
to update and maintain the frameworks, and it is suitable for industrial
applications in the fields of machine vision, multimedia, and speech. Caffe
works well in computer vision models compared to other domains in deep
learning [86].
Caffe can work with many different types of deep learning architectures.
The framework is suitable for various architectures such as CNN, LRCN, and
LSTM. A large number of pre-​configured training models are available to
the user, allowing a quick introduction to machine learning and the use of
neural networks [87]. As a platform for Caffe come Linux distributions such
as Ubuntu but also macOS and Docker container in question. For Windows
installations, solutions are also available on GitHub. For the Amazon AWS
Cloud, Caffe is available as a preconfigured Amazon Machine Image [88].

6.3.2.3 PyTorch
Nowadays, in the domain of deep learning, PyTorch is one of the popular
frameworks among researchers around the world. It is also an open-​source
framework developed by Facebook AI Research Group; it is a pythonic way
of implementing our deep learning models and it provides all the services
and functionalities offered by the Python environment. It allows auto differ-
entiation that helps to speed up the backpropagation process [89]. PyTorch
comes with various modules like torchvision, torchaudio, and torchtext,
which is flexible to work in neural language processing (NLP) and computer
Brain–Computer Interface for Dream Visualization 109

vision [90]. PyTorch is more flexible for the researcher than for developers.
It provides better performance compared to Keras and Caffe, but it is a low-​
level API that focused on direct work on array expressions [91].
PyTorch is a Python-​based scientific computing package that is a replace-
ment for NumPy to make use of the power of GPUs and a deep learning
research platform that provides maximum flexibility and speed [92]. It
ensures an easy-​to-​use API, which helps with easier usability and better
understanding when making use of the API [93]. It is fast and feels native,
hence ensuring easy coding and fast processing. The support for CUDA
ensures that the code can run on the GPU, thereby decreasing the time
needed to run the code and increasing the overall performance of the
system [94].

6.3.2.4 MATLAB
In the domain of electronics and communication, MATLAB is the most
popular software tool since early times. However, it is losing popularity
among modern engineers and researchers. A special Deep Learning Toolbox
is developed under the umbrella of MATLAB for processing and analyzing a
dataset by using a new model, pre-​trained models, and apps [95]. However,
it provides interfaces for designing and implementing most of the popular
neural network models such as CNN, RNN, GAN, and LSTM. Here CNN
and LSTM are used to perform classification and regression on image, time
series, and text data. Further, GAN and Siamese networks can also be used
for automatic differentiation, custom training loops, and shared weights [96].
It uses its programming language and environment, which is different from
the Python environment. This tool is originally used for signal processing.
However, models can also be imported and exchanged in the Python envir-
onment through other popular deep learning tools described in previous
sections (TensorFlow, Caffe, and PyTorch) into the MATLAB environment
by using the ONNX tool. The toolbox also supports transfer learning with
DarkNet-​53, ResNet-​50, NASNet, SqueezeNet, and many other pretrained
models [97].
Deep Network Designer and Experiment Manager are additional
applications with MATLAB Deep Learning Toolbox. Deep Network Designer
is helpful for designing, analyzing, and training neural networks graphically
[98,99]. The Experiment Manager helps in managing multiple deep learning
experiments, keeping track of training parameters, analyzing results, com-
paring code from different experiments, visualizing layer activations, and
graphically monitoring training progress [100]. It is easily possible to speed
up training using multiple GPU machines or scale up to clusters and clouds,
including NVIDIA-​GPU Cloud and Amazon EC2-​GPU instances [101].
Comparative Study of Keras, Caffe, PyTorch, MATLAB Deep Learning
Frameworks are shown in Table 6.4.
110 Machine Learning Algorithms and Applications in Engineering

TABLE 6.4
Comparative Study of Various Deep Learning Frameworks

MATLAB (Deep-​
Parameter Keras Caffe PyTorch Learning Toolbox)

Source Open source Open source Open source Commercial


API High level High-​level Low level Low-​Level
Language Python Python Python MATLAB
Developer Google Brain Berkeley University Facebook Math-​Works
User Beginners Engineers Researchers Engineers

6.4 Conclusion
This chapter provides just an overview of a dream communication system.
However, it’s not easy to implement such a type of system, but this is a step
toward such significant technology. It may be available among the people
in near future. It is considered that information can be processed inside the
brain in the same manner as the computer works, while both dreaming and
awake. Hence information processing theory is most significant among all
other theories of dreaming. Further, it can be seen that during the REM sleep
stage, mostly dreams occur in the mind. On the other hand, it can also be seen
that most of the memory consolidation happens during the REM sleep stage,
which means temporary memory is transferred from the hippocampus to
associated neocortical regions for permanent storage, which is also known as
long-​term memory. Dreaming is the result of memory consolidation, which
is a way of learning things by strengthening cortical connections of chemical
composition or neurotransmitters. It is possible to read and perform func-
tional activities while dreaming in the REM sleep stage by using various hard-
ware modalities frequently used in the domain of brain–​computer interface
technology. Finally, we have to map these signals or imageries using deep
learning-​based models clearly with having better audio and video qualities.

References
[1] Thomas Metzinger, Why are dreams interesting for philosophers? The example
of minimal phenomenal selfhood, plus an agenda for future research, Frontiers
in Psychology, 4:746, 2013.
[2] Walinga and Charles Stangor, Introduction to Psychology, BC-​ Campus Open
Education.
[3] Ricardo A. Velluti, Interactions between sleep and sensory physiology, J. Sleep
Res., 1997, 6, 61–​77.
Brain–Computer Interface for Dream Visualization 111

[4] Julian Mutz, Amir-​Homayoun Javadi, Exploring the neural correlates of dream
phenomenology and altered states of consciousness during sleep, Neuroscience
of Consciousness, Volume 2017, Issue 1, 2017, nix009, https://​doi.org/​10.1093/​
nc/​nix​009
[5] Gwen Dewar, Sleep Requirements, Parenting Science Store.
[6] Sander van der Linden, The science behind dreaming, Scientific American, 2011.
[7] Sigmund Freud, The Interpretation of Dreams, 1900, https://​psychc​lass​ics.yorku.
ca/​Freud/​Dre​ams/​dre​ams.pdf.
[8] Wei Zhang and Benyu Guo, Freud’s dream interpretation: A different perspec-
tive based on the self-​organization theory of dreaming, Front. Psychol., Vol 9,
2018, DOI=​10.3389/​fpsyg.2018.01553.
[9] Eugen Tarnow, How dreams and memory may be related, Neuropsychoanalysis, 2014.
[10] Chester A. Pearlman, REM sleep and information processing: Evidence from
animal studies, Neurosci. Biobehav. Rev., 1979, 57–​68.
[11] Erin J. Wamsley and Robert Stickgold, Dreaming and offline memory pro-
cessing, Curr. Biol., 2010 Dec 7;20(23):R1010-​3. doi: 10.1016/​j.cub.2010.10.045.
PMID: 21145013; PMCID: PMC3557787.
[12] Michael S. Franklin, The role of dreams in the evolution of the human mind,
Evolutionary Psychology, January 2005. doi:10.1177/​147470490500300106.
[13] J. Allan Hobson, Robert W. McCarley, The brain as a dream state generator: An
activation-​synthesis hypothesis of the dream process, The American Journal of
Psychiatry, 1977.
[14] Eiser AS. Physiology and psychology of dreams. Semin Neurol. 2005
Mar;25(1):97-​105. doi: 10.1055/​s-​2005-​867078. PMID: 15798942.
[15] Lin Edwards, Dreams may have an important physiological function, Medical
Xpress, 2009.
[16] Gent TC, Bandarabadi M, Herrera CG, Adamantidis AR. Thalamic dual control
of sleep and wakefulness. Nat Neurosci. 2018 Jul;21(7):974–​984. doi: 10.1038/​
s41593-​018-​0164-​7. Epub 2018 Jun 11. PMID: 29892048; PMCID: PMC6438460.
[17] Brown RE, Basheer R, McKenna JT, Strecker RE, McCarley RW. Control of
sleep and wakefulness. Physiol Rev. 2012 Jul;92(3):1087–​ 187. doi: 10.1152/​
physrev.00032.2011. PMID: 22811426; PMCID: PMC3621793.
[18] Lukas T. Oesch, Mary Gazea, Thomas C. Gent, Mojtaba Bandarabadi, Carolina
Gutierrez Herrera, Antoine R. Adamantidis, REM sleep stabilizes hypothal-
amic representation of feeding behavior, Proceedings of the National Academy of
Sciences, 117 (32), 19590–​19598, 2020, https://​doi.org/​10.1073/​pnas.192190​911.
[19] Corsi-​ Cabrera M, Velasco F, Del Río-​ Portilla Y, Armony JL, Trejo-​ Martínez
D, Guevara MA, Velasco AL. Human amygdala activation during rapid eye
movements of rapid eye movement sleep: an intracranial study. J Sleep Res. 2016
Oct;25(5):576–​582. doi: 10.1111/​jsr.12415. Epub 2016 May 5. PMID: 27146713.
[20] Brain Basics: Understanding Sleep. www.ninds.nih.gov/​Disord​ers/​Pati​ent-​
Caregi​ver-​Educat​ion/​Unders​tand​ing-​Sleep.
[21] Sleep Physiology, National Academies Press, 2006.
[22] Memories involve replay of neural firing patterns. www.nih.gov/​news-​eve​nts/​
nih-​resea​rch-​matt​ers/​memor​ies-​invo​lve-​rep​lay-​neu​ral-​fir​ing-​patte​rns.
[23] Rasch B, Born J. About sleep's role in memory. Physiol Rev. 2013;93(2):681–​766.
doi:10.1152/​physrev.00032.2012
[24] Takeuchi Tomonori, Duszkiewicz Adrian J. and Morris Richard G. M. 2014.
The synaptic plasticity and memory hypothesis: encoding, storage and
112 Machine Learning Algorithms and Applications in Engineering

persistence, Phil. Trans. R. Soc. B369:2013028820130288 http://​doi.org/​10.1098/​


rstb.2013.0288.
[25] Pisella L, Alahyane N, Blangero A, Thery F, Blanc S, Pelisson D. Right-​
hemispheric dominance for visual remapping in humans. Philos Trans R Soc Lond
B Biol Sci. 2011;366(1564):572-​585. doi:10.1098/​rstb.2010.0258.
[26] Helder Bértolo, Visual imagery without visual perception? Psicológica 26 (2005)
173–​188.
[27] Chai M. Tyng, Hafeez U. Amin, Mohamad N. M. Saad, and Aamir S. Malik, The
influences of emotion on learning and memory, Front Psychol. 8 (2017) 1454.
[28] Marcelo R. Roxo, Paulo R. Franceschini, Carlos Zubaran, Fabrício D. Kleber, and
Josemir W. Sander, The limbic system conception and its historical evolution,
Scientific World Journal 11 (2011) 2428–​2441.
[29] C. Daniel Salzman, Amygdala. www.bri​tann​ica.com/​scie​nce/​amygd​ala
[30] Serena Scarpelli, Chiara Bartolacci, Aurora D’Atri, Maurizio Gorgoni, and Luigi
De Gennaro, The functional role of dreaming in emotional processes. Front
Psychol. 10 (2019) 459.
[31] Albert McKeon, Oneirology is helping clarify our understanding of dreams,
https://​now.nort​hrop​grum​man.com/​one​irol​ogy-​is-​help​ing-​clar​ify-​our-​unders​
tand​ing-​of-​dre​ams/​
[32] Guangye Li and Dingguo Zhang, Brain–​ computer interface controlled
cyborg: establishing a functional information transfer pathway from human
brain to cockroach brain, PLoS One 11(3) (2016) e0150667.
[33] Joseph N. Mak and Jonathan R. Wolpaw, Clinical applications of brain–​computer
interfaces: Current state and future prospects. IEEE Rev Biomed Eng. 2 (2009)
187–​199.
[34] Scott Makeig, Christian Kothe, and Tim Roger Mullen, Evolving signal pro-
cessing for brain–​ computer interfaces, Proceedings of the IEEE 100(Special
Centennial Issue): 1567–​1584.
[35] Claude Frasson and George Kostopoulos, Brain function assessment in learning.
First International Conference, BFAL 2017, Patras, Greece, September 24–​25,
2017, Proceedings.
[36] Jason Brownlee, What is deep learning? https://​mac​hine​lear​ning​mast​ery.com/​
what-​is-​deep-​learn​ing/​
[37] Lopes da Silva, Fernando. (2010). EEG: Origin and measurement. 10.1007/​978-​
3-​540-​87919-​0_​2.
[38] Jessica D. Payne and Lynn Nadel, Sleep, dreams, and memory consolidation: The
role of the stress hormone cortisol, Learning Memory, 11(6) ( 2004) 671–​678.
[39] Ian G. Campbell, EEG recording and analysis for sleep research, Current Protocols
in Neuroscience. 2009 October. Chapter: Unit 10.2.
[40] What is an electrocardiogram (ECG)? Institute for Quality and Efficiency in
Health Care (IQWiG), 2006.
[41] Kendra Cherry, Why do we dream? www.veryw​ellm​ind.com/​why-​do-​we-​
dream-​top-​dream-​theor​ies-​2795​931
[42] Sara Reardon, Dream movements translate to real life, www.sci​ence​mag.org/​
news/​2011/​10/​dream-​moveme​nts-​transl​ate-​real-​life
[43] Magnetic resonance imaging (MRI). www.nibib.nih.gov/​scie​nce-​educat​ion/​
scie​nce-​top​ics/​magne​tic-​resona​nce-​imag​ing-​mri
[44] Kathryn Mary Broadhouse, The physics of MRI and how we use it to reveal the mys-
teries of the mind. https://​kids.fron​tier​sin.org/​arti​cle/​10.3389/​frym.2019.00023
Brain–Computer Interface for Dream Visualization 113

[45] Magnetic resonance imaging overview, www.pauldu​rso.com/​MRI​_​tec​hnol​


ogy.htm
[46] Magnetic resonance imaging (MRI), www.cdt-​babes.ro/​medi​cal-​artic​les/​
magne​tic-​resona​nce-​imag​ing-​mri.php
[47] The brain may actively forget during dream sleep, www.nih.gov/​news-​eve​nts/​
news-​relea​ses/​brain-​may-​activ​ely-​for​get-​dur​ing-​dream-​sleep
[48] Singh SP. Magnetoencephalography: Basic principles. Ann Indian Acad Neurol.
2014;17(Suppl 1):S107–​S112. doi:10.4103/​0972-​2327.128676.
[49] What is a MEG scanner? www.chop.edu/​tre​atme​nts/​mag​neto​ence​phal​ogra​
phy-​meg-​scan.
[50] Hamza Alizai, Gregory Chang, and Ravinder R. Regatte. Sample records for
field MRI systems, www.scie​nce.gov/​top​icpa​ges/​f/​field+​mri+​syst​ems.
[51] Luis Fernando Nicolas-​Alonso, Jaime Gomez-​Gil, Brain–​computer interfaces, a
review. Sensors (Basel) 12(2) (2012) 1211–​1279; Jerry J. Shih, Dean J. Krusienski,
Jonathan R. Wolpaw, Mayo Clin Proceedings 87(3) (2012) 268–​279.
[52] Dahlia Sharon, Matti S Hämäläinen, Roger B.H. Tootell, Eric Halgren, John W
Belliveau. The advantage of combining MEG and EEG: comparison to fMRI in
the focally-​stimulated visual cortex, Neuroimage 36(4) (2007) 1225–​1235.
[53] Gregory A. Light, Lisa E. Williams, Falk Minow, Joyce Sprock, Anthony Rissling,
Richard Sharp, Neal R. Swerdlow, and David L. Braff, Electroencephalography
(EEG) and event-​related potentials (ERPs) with human participants. Current
Protocols in Neuroscience. 2010 Jul; Chapter: Unit–​6.2524.
[54] Hafeez Ullah Amin, Wajid Mumtaz, Ahmad Rauf Subhani, Mohamad Naufal,
and Mohamad Saad, Classification of EEG signals based on pattern recognition
approach, November 2017. Frontiers in Computational Neuroscience, https://​doi.
org/​10.3389/​fncom.2017.00103
[55] Hong Liang, Xiao Sun, Yunlei Sun, Yuan Gao, Text feature extraction based
on deep learning: a review, EURASIP Journal on Wireless Communications and
Networking, Volume 2017.
[56] Jason Brownlee, https://​mac​hine​lear​ning​mast​ery.com/​an-​intro​duct​ion-​to-​feat​
ure-​select​ion/​
[57] Anil K. Jain, Robert Duin, and Jianchang Mao, Statistical pattern recogni-
tion: A review, IEEE Transactions on Pattern Analysis and Machine Intelligence 22(1)
(2000) 4–​37.
[58] Shahadat Uddin, Arif Khan, Md Ekramul Hossain, and Mohammad Ali Moni,
Comparing different supervised machine learning algorithms for disease pre-
diction, BMC Medical Information Decision Making 19 (2019) 281.
[59] Human–​computer interaction, https://​psu.pb.uni​zin.org/​ist​110/​chap​ter/​5-​2-​
human-​compu​ter-​inte​ract​ion/​
[60] Swati Aggarwal and Nupur Chugh, Signal processing techniques for motor
imagery brain–​ computer interface: A review, Array, Volumes 1–​ 2, January–​
April 2019.
[61] V. Sharma, Machine learning: Introduction to regression analysis, https://​
vin​odsb​log.com/​2019/​02/​01/​mach​ine-​learn​ing-​intro​duct​ion-​to-​reg​ress​ion-​
analy​sis/​
[62] Tim Urban, The AI revolution: Our immortality or extinction, https://​wai​tbut​
why.com/​2015/​01/​art​ific​ial-​intel​lige​nce-​rev​olut​ion-​2.html
[63] Cesar Timo-​Iaria (in memoriam); Angela Cristina do Valle, Physiology of dreaming,
https://​sleep​scie​nce.org.br/​deta​ils/​145/​en-​US/​phy​siol​ogy-​of-​dream​ing
114 Machine Learning Algorithms and Applications in Engineering

[64] Connor Shorten, Taghi M. Khoshgoftaar, A survey on image data augmen-


tation for deep learning, https://​link.sprin​ger.com/​arti​cle/​10.1186/​s40​
537-​019-​0197-​0
[65] Deep learning algorithms you should know about, www.simp​lile​arn.com/​
deep-​learn​ing-​alg​orit​hms-​arti​cle
[66] Vibhor Nigam, Understanding neural networks. From neuron to RNN, CNN,
and deep learning, https://​tow​ards​data​scie​nce.com/​unders​tand​ing-​neu​ral-​
netwo​rks-​from-​neu​ron-​to-​rnn-​cnn-​and-​deep-​learn​ing-​cd88e​90e0​a90
[67] Geoffrey E. Hinton, Simon Osindero, Yee-​Whye Teh; A Fast Learning Algorithm
for Deep Belief Nets. Neural Comput 2006; 18 (7): 1527–​1554. doi: https://​doi.
org/​10.1162/​neco.2006.18.7.1527.
[68] Deep belief networks: How they work and what are their applications, https://​
miss​ingl​ink.ai/​gui​des/​neu​ral-​netw​ork-​conce​pts/​deep-​bel​ief-​netwo​rks-​work-​
appli​cati​ons/​
[69] Jason Brownlee, A gentle introduction to RNN unrolling, https://​mac​hine​lear​
ning​mast​ery.com/​rnn-​unroll​ing/​
[70] Introduction to recurrent neural network, www.geeksf​orge​eks.org/​intro​duct​
ion-​to-​recurr​ent-​neu​ral-​netw​ork/​
[71] Mahendran Venkatachalam, Recurrent neural networks, https://​tow​ards​data​
scie​nce.com/​recurr​ent-​neu​ral-​netwo​rks-​d4642​c9bc​7ce
[72] V. Sharma, Deep learning: Introduction to recurrent neural networks, https://​
vin​odsb​log.com/​2019/​01/​07/​deep-​learn​ing-​intro​duct​ion-​to-​recurr​ent-​neu​ral-​
netwo​rks/​
[73] Sumit Saha, A comprehensive guide to convolutional neural networks, https://​
tow​ards​data​scie​nce.com/​a-​compre​hens​ive-​guide-​to-​convol​utio​nal-​neu​ral-​
netwo​rks
[74] Shadab Hussain, Building a convolutional neural network, https://​tow​ards​
data​scie​nce.com/​build​ing-​a-​convol​utio​nal-​neu​ral-​netw​ork-​male-​vs-​fem​ale
[75] Convolutional neural network (ConvNet/​ CNN), www.sel​fdri​ving​cars​360.
com/​gloss​ary/​convol​utio​nal-​neu​ral-​netw​ork/​
[76] Jason Brownlee, A gentle introduction to generative adversarial networks
(GANs), https://​mac​hine​lear​ning​mast​ery.com/​what-​are-​gen​erat​ive-​adve​rsar​
ial-​netwo​rks-​gans/​
[77] The discriminator, https://​dev​elop​ers.goo​gle.com/​mach​ine-​learn​ing/​gan/​
discri​mina​tor
[78] Lee JG, Jun S, Cho YW, Lee H, Kim GB, Seo JB, Kim N. Deep Learning in
Medical Imaging: General Overview. Korean J Radiol. 2017 Jul-​Aug;18(4):570–​
584. doi: 10.3348/​kjr.2017.18.4.570. Epub 2017 May 19. PMID: 28670152;
PMCID: PMC5447633.
[79] Kate Brush and Ed Burns, Deep learning, https://​sea​rche​nter​pris​eai.tec​htar​get.
com/​def​i nit​ion/​deep-​learn​ing-​deep-​neu​ral-​netw​ork
[80] Ketkar, Nikhil. (2017). Introduction to Keras. 10.1007/​978-​1-​4842-​2766-​4_​7.
[81] Purva Huilgol, Pre-​trained models for image classification with Python code,
www.anal ​ y tic ​ s vid ​ h ya.com/ ​ b log/ ​ 2 020/ ​ 0 8/ ​ t op- ​ 4 - ​ p re- ​ t rai ​ n ed- ​ m od ​ e ls- ​ f or-​
image-​cla​ssif​i cat​ion-​with-​pyt​hon-​code/​
[82] Gilbert Tanner, Introduction to deep learning with keras, https://​tow​ards​data​
scie​nce.com/​intro​duct​ion-​to-​deep-​learn​ing-​with-​keras
Brain–Computer Interface for Dream Visualization 115

[83] Martin Heller, What is keras? The deep neural network API explained, www.
infowo​rld.com/​arti​cle/​3336​192/​what-​is-​keras-​the-​deep-​neu​ral-​netw​ork-​api-​
explai​ned.html
[84] Why use keras? https://​mran.micros​oft.com/​snaps​hot/​2018-​01-​07/​web/​
packa​ges/​keras/​vignet​tes/​why_​us​e_​ke​ras.html
[85] Gautam Ramuvel, Best open source frameworks for machine learning, https://​
med​ium.com/​coinmo​nks/​5-​best-​open-​sou​rce-​fra​mewo​rks-​for-​mach​ine-​learn​
ing-​739d0​6170​601
[86] What is caffe—​The deep learning framework, https://​cod​ingc​ompi​ler.com/​
what-​is-​caffe/​
[87] Evan Shelhamer, Deep learning for computer vision with caffe and cuDNN, https://​
develo​per.nvi​dia.com/​blog/​deep-​learn​ing-​compu​ter-​vis​ion-​caffe-​cudnn/​
[88] Martin Heller, Caffe: deep learning conquers image classification, www.infowo​
rld.com/​arti​cle/​3154​273/​rev​iew-​caffe-​deep-​learn​ing-​conqu​ers-​image-​cla​ssif​
icat​ion.html
[89] Caffe Python 3.6 NVidia GPU Production on Ubuntu, https://​aws.ama​
zon.com/​mark​etpl​ace/​pp/​Jetw​are-​Caffe-​Pyt​hon-​36-​NVi​dia-​GPU-​Pro​duct​
ion-​on-​U/​
[90] Keras vs PyTorch vs Caffe: Comparing the implementation of CNN, https://​
analyt​icsi​ndia​mag.com/​keras-​vs-​pyto​rch-​vs-​caffe-​compar​ing-​the-​imp​leme​
ntat​ion-​of-​cnn/​
[91] Deep learning frameworks compared: Keras vs PyTorch vs Caffee, https://​cont​
ent.tech​gig.com/​deep-​learn​ing-​fra​mewo​rks-​compa​red-​keras-​vs-​pyto​rch-​vs-​
caf​fee/​arti​cles​how/​77480​133.cms
[92] Tensorflow vs Keras vs Pytorch: Which framework is the best? https://​med​
ium.com/​@Atlas​Syst​ems/​ten​sorf​l ow-​vs-​keras-​vs-​pyto​rch-​which-​framew​ork-​
is-​the-​best-​f92f9​5e11​502
[93] Implementing deep neural networks using PyTorch, https://​med​ium.com/​
edur​eka/​pyto​rch-​tutor​ial-​9971d​66f6​893
[94] Deploying Pytorch in Python via a Rest API with Flask, https://​pyto​rch.org/​
tutori​als/​inter​medi​ate/​flas​k_​re​st_​a​pi_​t​utor​ial.html
[95] Cuda Semantics, https://​pyto​rch.org/​docs/​sta​ble/​notes/​cuda.html
[96] Deep Learning Toolbox, www.mathwo​rks.com/​produ​cts/​deep-​learn​ing.html
[97] Introducing deep learning with MATLAB, https://​in.mathwo​rks.com/​campai​
gns/​off​ers/​deep-​learn​ing-​with-​mat​lab.html
[98] Pretrained deep neural networks, https://​in.mathwo​rks.com/​help/​deepl​earn​
ing/​ug/​pre​trai​ned-​convol​utio​nal-​neu​ral-​netwo​rks.html
[99] Build networks with deep network designer, www.mathwo​rks.com/​help/​
deepl​earn​ing/​ug/​build-​netwo​rks-​with-​deep-​netw​ork-​desig​ner.html
[100] Experiment manager, www.mathwo​rks.com/​help/​deepl​earn​ing/​ref/​experi​
ment​mana​ger/​
[101] Cloud and data centre, www.nvi​dia.com/​en-​in/​data-​cen​ter/​gpu-​cloud-​
comput​ing/​
newgenprepdf

7
Machine Learning and Data Analysis
Based Breast Cancer Classification

Souvik Das,* Rama Chaitanya Karanam, Obilisetty Bala Krishna, and


Jhareswar Maiti
IIT Kharagpur, Kharagpur, West Bengal
E-​mail: *[email protected]

CONTENTS
7.1 Introduction..................................................................................................117
7.1.1 Review of the Literature.................................................................118
7.2 Methodology................................................................................................118
7.2.1 Evaluation of Machine Learning Models.................................... 122
7.2.2 Data Description............................................................................. 122
7.3 Heat Map of Correlation........................................................................... 123
7.4 Results and Discussions............................................................................ 123
7.5 Conclusion................................................................................................... 128

7.1 Introduction
The current study preprocesses the cancer patient data, and machine learning
models will be trained with this data to classify breast cancer. The study
mainly consists of five parts: (i) data collection and understanding, (ii) data
preprocessing, (iii) training the ML models, (iv) evaluation, and (v) deploy-
ment. We have extracted the required data from the images of cancer cells
of 570 patients. We took 30 attributes from the images to classify cancer. The
features are computed from a digitized image of a fine needle aspirate (FNA)
of a breast mass. They describe the characteristics of the cell nuclei present
in the image. For applying machine learning techniques on the data set, we
need to preprocess and dispute the data to get a better data set. For analyzing
the data set, we need specific Python libraries and packages like Pandas for
data manipulation and analysis, NumPy for mathematical operations on the

DOI: 10.1201/9781003104858-7 117


118 Machine Learning Algorithms and Applications in Engineering

data set, Matplotlib for data visualization, Seaborn for making heatmaps, and
Sklearn package for normalization of the data set. In data preprocessing, we
remove unnecessary columns from the data frame, change categorical values
to numerical values, and standardize the data.
Machine learning is often used to train machines on how to handle the
data more efficiently. Here, we have applied supervised learning algorithms
to train the model. These include KNN, decision tree, logistic regression,
and support vector machine (SVM). The main advantage of using machine
learning is that, once an algorithm learns what to do with data, it can do its
work automatically. Then we evaluate these models using the Metrics library
of Sklearn package. We select the model with high accuracy and deploy it to
classify the cancer of the new patient using his cancer cells data.

7.1.1 Review of the Literature


In the literature, it is found that artificial intelligence and Internet of Things
(IoT)-​based techniques are used to improve the accuracy of the diagnosis of
the breast cancer and its prediction. We have provided some of the state-​of-​
the-​art machine learning and deep learning approaches, which are used to
solve the problem of breast cancer.
Arpit and Aruna [1] optimized a neural network architecture by applying
new crossover and mutants and proposed a genetically optimized neural net-
work (GONN) to classify breast cancer.
Ibrahim and Samsuddin [2] proposed a multilayer perception-​based neural
network model for classifying breast cancer accurately and precisely.
Liu et al. [3] proposed simulated annealing based genetic algorithm and
cost-​sensitive SVM to improve the classification model for breast cancer
diagnosis.
Zemmal [4] presented a computer-​assisted genetic algorithm integrated
semi-​ supervised SVM to propose an intelligent breast cancer classifica-
tion model.
Helwan [5] developed an automated system for the classification of breast
cancer based on feed-​forward neural network using the back propagation
and radial basis function algorithm.
Based on the above-​mentioned literature, it can be inferred that more
sophisticated machine learning models are to be used for the classification
of breast cancer.

7.2 Methodology
Here, we describe a few concepts and algorithms, which we have used for
constructing the machine learning models for breast cancer classification.
ML in Breast Cancer Classification 119

(i) Data preprocessing: The raw data is highly vulnerable to missing data,
outliers, noise, and inconsistent data because of the vastness of data, mul-
tiple resources, and their gathering methods [6]. The poor quality of data
profoundly affects the results of machine learning models. Therefore,
preprocessing techniques must be applied to the data to get better results
of ML models [7]. After preprocessing, the data must be transformed into
the required form for training the ML models [8]. Data preprocessing
methods are data cleaning, filling missing values, removing outliers from
data, and standardization of data [9]. Missing values of a column data
can be filled with the mean of the column data. Outliers can be removed
using the binning process. Standardization of data gives a better classi-
fication because all the features will be constricted in the range of [–​1,1]
[10]. We should split the data to train and test data sets in the ratio of 7:3.
  Train data set will be used to train the ML model and test data set is
used to evaluate the trained ML model [11]. Cross-​validation techniques
belong to conventional approaches used to ensure good generalization
and to avoid overtraining. The basic idea is to divide the data set T
into two subsets—​one subset is used for training while the other sub-
group is left out, and the performance of the final model is evaluated
on it. The primary purpose of cross-​validation is to achieve a stable and
confident estimate of the model performance [12]. Cross-​validation
techniques can also be used when evaluating and mutually comparing
more models, various training algorithms, or when seeking optimal
model parameters [13].
(ii) Training of ML models: The purpose of machine learning is to learn from
the data. Many studies have been done on how to make machines learn
by themselves. Many mathematicians and programmers apply several
approaches to find a solution to this problem. Some of them are KNN,
logistic regression, decision tree, SVM, and random forest. Logistic
regression is a classification function that uses a class for building and
uses a single multinomial logistic regression model with a single esti-
mator [14]. SVM is the most recent supervised machine learning tech-
nique that revolves around the notion of a margin on either side of a
hyperplane that separates two data classes. Maximizing the margin
and thereby creating the most significant possible distance between the
separating hyperplane and the instances on either side of it has been
proven to reduce an upper bound on the expected generalization error
[14]. Decision trees are those types of trees that groups attribute by
sorting them based on their values. The decision tree is used mainly for
classification purposes. Each tree consists of nodes and branches. Each
node represents attributes in a group that is to be classified, and each
branch represents a value that the node can take. In k-​NN classification,
the output is a class membership. An object is classified by a plurality
vote of its neighbors, with the object being assigned to the class most
common among its k nearest neighbors [14].
120 Machine Learning Algorithms and Applications in Engineering

(iii) Machine learning algorithms: Machine learning algorithms, which are


used in this study, are explained below.
A. K-​Nearest neighbors (KNN): It calculates the distance between the
test data point and all the training data points using Euclidean dis-
tance formula to check the K nearest neighbors’ classes and mode of
the k nearest classes will be returned.

Euclidean distance: (X1 − Y1 )2 + (X 2 − Y2 )2 +  + (X n − Yn )2

B. Linear regression: It models a relationship between independent


variables and the dependent variable (predictor variable) using a
gradient descent approach. After modeling, it can be used to pre-
dict the value for a given data.
^
Let’s suppose X1, X2, … Xn are the
independent variables and Y is the predicted value of a dependent
variable. Linear regression calculates the weights for each inde-
pendent variable and returns an equation.

^
Y = α + β1 + X1 + β 2 + X 2 +  + β n + X n

The weights will be calculated using cost function and gradient


descent approach.
Cost function:

^ 1 ^
cos t(Y , Y ) = (Y , Y )2
2

1 n

^
J (β) = cos t(Y i , Yi )
n i =1

This J(β) function will be plotted in a (n+​1) dimensional plane to


find the local minimum, that is, to find the values of the weights at
which the cost function is minimum. These values will be used for
predictions.

C. Logistic regression: It uses the sigmoid function to reduce the range


of values predicted in linear regression to [0,1]. The output of the
sigmoid function is the probability of a data classifying the class
mapped to value 1.

^ 1
Sigmoid function: σ(Y ) =
1 + e −Y
^
Probability that the output is 1: p(Y = ​1| X) =​ σ(Y )
^
Probability that the output is 0: p(Y = ​0|X) = 1 − σ(Y )
ML in Breast Cancer Classification 121

^
If σ(Y ) is greater than or equals to 0.5, logistic regression returns 1 as
output, else returns 0 as output.

D. Decision tree: A decision tree is a flowchart-​like structure in which each


internal node represents a “test” on an attribute (e.g., whether a coin
flip comes up heads or tails), each branch represents the outcome of
the test, and each leaf node represents a class label (decision taken after
computing all attributes). The selection of an attribute is based on the
entropy of the data and difference in the entropy after selecting an
attribute.

Entropy: The randomness of the data set. Here p is the possibility of


selecting a class while being randomly picked from a data set:

Entropy =​ − p(log( p)) − (1 − p)(log(1 − p))

Information gain: The information gain is based on the decrease in entropy


after a data set is split on an attribute. Constructing a decision tree is all
about finding attribute that returns the highest information gain.

Information gain =​E1 –​ E2

E1 is entropy of data before splitting.

E2 is entropy of data after splitting.

E. Support Vector Machines: A SVM is a discriminative classifier formally


defined by a separating hyperplane. In other words, given labeled
training data (supervised learning), the algorithm outputs an optimal
hyperplane, which categorizes new examples. In two-​ dimensional
space, this hyperplane is a line dividing a plane in two parts where in
each class lay on either side.

SVM uses different kernel functions to map the data into high dimen-
sional plane, where the data can be separated linearly.
Kernel functions used in SVM are the following:

1.  Polynomial function


2.  Sigmoid function
3.  Radial basis function

N
y( x ) = ∑ wi ϕ( x − xi )
i =1
122 Machine Learning Algorithms and Applications in Engineering

F. Random forest classification: It creates n different data sets from the


training data set and creates a decision tree for each data set. N can
be changed manually. Each decision tree will classify the class for the
given test data and the mode of the results of n decision trees will be
given as output.

7.2.1 Evaluation of Machine Learning Models


The metrics we can use to evaluate a machine learning model are accuracy,
confusion matrix, logarithmic loss, mean absolute error, mean squared
error, precision, and recall. Accuracy is defined as the percentage of correct
predictions for the test data. It can be calculated easily by dividing the
number of correct classifications by the number of total numbers of testing
data. Precision is defined as the fraction of relevant examples (true positives)
among all the examples that were predicted to belong to a certain class. The
recall is defined as the fraction of examples that were predicted to belong
to a class with respect to all of the examples that truly belong in the class.
The confusion matrix returns us a matrix, which can be used to describe the
performance of a model. The confusion matrix gives the best results in the
case of two-​class classification problems. The logarithmic loss will be more
useful in the case of multiclass classification problems. Accuracy works well
only if there are an equal number of samples for each class. F1 score gives a
balance between precision and recall. F1 score also can be used to evaluate
ml models.

7.2.2 Data Description
The data that is used in the current study for classifying breast cancer contains
data of 570 patients of Wisconsin hospital with 32 attributes. The attributes
information is given below.
ID number, and Diagnosis (M =​malignant, B =​benign) are the general
attributes.
Ten real-​valued attributes are considered and given as follows: (a) radius
(mean of distances from center to points on the perimeter), (b) texture (standard
deviation of gray-​scale values), (c) perimeter, (d) area, (e) smoothness (local
variation in radius lengths), (f) compactness (perimeter^2 /​area –​1.0),
(g) concavity (severity of concave portions of the contour), (h) concave points
(number of concave portions of the contour), (i) symmetry, and (j) fractal
dimension (“coastline approximation” –​1)
The mean, standard error, and “worst” or largest (mean of the three lar-
gest values) of these features were computed for each image, resulting in 30
features.
ML in Breast Cancer Classification 123

FIGURE 7.1
Correlational plot.

7.3 Heat Map of Correlation


After analyzing the heat map (see Figure 7.1), we found that radius mean,
perimeter mean, and area mean are highly correlated. From these, we select
one feature for further analysis. We also noted that concave points mean, con-
cavity mean, and compactness mean area highly correlated. In a similar way,
we select one feature from these and thus we have removed unnecessary
columns and created a new data set.

7.4 Results and Discussions


After preprocessing the raw data, we have applied previously discussed
machine learning algorithms on the preprocessed data to classify breast
cancer. Models were created and trained in the Jupyter notebook environment
and code was written in Python language. The required packages have been
installed and used to ease the work. Accuracy score, confusion matrix, preci-
sion score, recall score, and f1 score were calculated using Sklearn package.
The results are discussed below. Figures 7.2–​7.9 provide the variations in
accuracy with respect to K-​fold for different machine learning algorithms.

(i)  KNN
(a) The accuracy of this model is 93 percent with a k value of 7.
(b) Maximum K-​fold cross-​validation accuracy is 93.75 percent with a
k value of 5.

The F1 score of the KNN algorithm is 92.95 percent.


124 Machine Learning Algorithms and Applications in Engineering

FIGURE 7.2
Accuracy graph for KNN.

FIGURE 7.3
Cross validation accuracy graph for KNN.

(ii) Decision Tree
The accuracy is 89.9 percent. Maximum K-​fold cross-​validation accuracy is
93.6 percent with a k value of 14.
The F1 score of decision tree algorithm is 92.61 percent.
ML in Breast Cancer Classification 125

FIGURE 7.4
Accuracy graph for Decision tree.

FIGURE 7.5
Accuracy graph for Logistic regression.
126 Machine Learning Algorithms and Applications in Engineering

(iii) Logistic Regression
The accuracy is 92.9 percent. Maximum K-​fold cross-​validation accuracy is
93.77 percent with a k value of 5.
The F1 score of logistic regression algorithm is 90.5 percent.

FIGURE 7.6
Accuracy graph for SVM (Rbf).

FIGURE 7.7
Accuracy graph for SVM (poly).
ML in Breast Cancer Classification 127

FIGURE 7.8
Accuracy graph for SVM (sigmoid).

FIGURE 7.9
Accuracy graph for Random Forest.

(iv) SVM (Rbf)
The accuracy is 92.89 percent. Maximum K-​fold cross-​validation accuracy is
95 percent with a k value of 12.
The F1 score of SVM (Rbf) algorithm is 90.78 percent.
128 Machine Learning Algorithms and Applications in Engineering

(v) SVM(Poly)
The accuracy is 89.89 percent. Maximum K-​fold cross-​validation accuracy is
91.37 percent with a k value of 12.
The F1 score of SVM (Poly) algorithm is 84.61 percent.

(vi) SVM(Sigmoid)
The accuracy is 88.16 percent. Maximum K-​fold cross-​validation accuracy is
88.07 percent with a k value of 7 or 8.
The F1 score of SVM (Sigmoid) algorithm is 84.5 percent.

(vii) Random Forest
The accuracy is 93.49 percent. Maximum K-​fold cross-​validation accuracy is
96.08 percent with a k value of 9.
The F1 score of random forest algorithm is 93.5 percent.
From the above results, we can conclude that random forest model is the
best model so far with a 96.08 percent accuracy and 0.935 F1 score.

7.5 Conclusion
The current study attempts to solve the problem of classifying breast
cancer based on the data obtained from raw images of breast cancer. We
have considered various state-​of-​art machine learning algorithms, such as
K-​nearest neighbor, decision tree, support vector machine, and random
forest. It is found that random forest is the best model for classifying breast
cancer. All the experiments and analysis are performed using Python library.
This study helps in formulating the process of diagnosing the breast cancer
from the preprocessed data obtained from raw images of breast cancer.
In future study, deep learning-​based tools and techniques can be used to
improve the prediction power and make the model more efficient. Future
selection process can also be added with the existing study to improve the
prediction accuracy.
ML in Breast Cancer Classification 129

References
[1] A. Bhardwaj and A. Tiwari, “Breast Cancer Diagnosis Using Genetically
Optimized Neural Network Model,” Expert Syst. Appl., pp. 1–​15, 2015.
[2] A. O. Ibrahim and S. M. Shansuddin, “Intelligent breast cancer diagnosis based
on enhanced Pareto optimal and multilayer perceptron neural network,” Int.
J. Comput. Aided Eng. Technol., vol. 10, no. 5, 2018.
[3] N. Liu, E. Qi, M. Xu, B. Gao, and G. Liu, “A novel intelligent classification model
for breast cancer diagnosis,” Inf. Process. Manag., vol. 56, no. 3, pp. 609–​623, 2019.
[4] N. Zemmal, N. Azizi, N. Dey, and M. Sellami, “Adaptive semi-​supervised
support vector machine semi supervised learning with features cooperation
for breast cancer classification,” J. Med. Imaging Heal. Informatics, vol. 6, no. 1,
pp. 53–​62, 2016.
[5] A. Helwan, J. B. Idoko, and R. H. Abiyev, “Machine learning techniques for clas-
sification of breast tissue,” Procedia Comput. Sci., vol. 120, no. 2017, pp. 402–​410,
2018, doi: 10.1016/​j.procs.2017.11.256.
[6] M. Han, Jiawei and Jian Pei, and Kamber, Data Mining: Concepts and Techniques.
Elsevier, Amsterdam, 2011.
[7] J. Han, “Proceedings of the 1996 ACM SIGMOD international conference on
management of data,” in Data Mining Techniques, 1996, p. 545.
[8] Soukup, Tom, and Davidson, Lan, Visual Data Mining: Techniques and Tools for
Data Visualization and Mining. John Wiley & Sons, 2002.
[9] G. Wang, Jianyong, and Karypis, “On efficiently summarizing categorical
databases,” Knowl. Inf. Syst., vol. 9, no. 1, pp. 19–​37, 2006.
[10] W. S. Alasadi, A. Suad, and Bhaya, “Review of data preprocessing techniques in
data mining,” J. Eng. Appl. Sci., vol. 12, no. 16, pp. 4102–​4107, 2017.
[11] H.R. Bowden, J. Gavin, and Graeme C. Dandy, and Maier, “Input determination
for neural network models in water resources applications. Part 1—​background
and methodology,” J. Hydrol., vol. 301, no. 1–​4, pp. 75–​92, 2005.
[12] W. G. Cochran, Sampling Techniques. John Wiley & Sons, 2007.
[13] D. Fernandes, Stenio, Carlos Kamienski, Judith Kelner, and Denio Mariz, and
Sadok, “A stratified traffic sampling methodology for seeing the big picture,”
Comput. Networks, vol. 52, no. 14, pp. 2677–​2689, 2008.
[14] J. Osisanwo, J.E.T. Akinsola, O. Awodele, J.O. Hinmikaiye, O. Olakanmi, and
Akinjobi, “Supervised machine learning algorithms: classification and com-
parison,” Int. J. Comput. Trends Technol., vol. 48, no. 3, pp. 128–​138, 2017.
newgenprepdf

8
Accurate Automatic Functional
Recognition of Proteins: Overview and
Current Computational Challenges

Javier Pérez-​Rodríguez,1,* Morteza Yazdani,2 and Prasenjit Chatterjee3


1
Department of Quantitative Methods, Universidad Loyola Andalucía, Córdoba, Spain
2
ESIC Business & Marketing School, Madrid, Spain
3
Department of Mechanical Engineering, MCKV Institute of Engineering, Howrah, India
*[email protected]

CONTENTS
8.1 Biological Framework................................................................................ 131
8.2 Identifying the Protein Functions............................................................ 133
8.3 Automatic Functional Annotation of Proteins....................................... 134
8.4 Challenges................................................................................................... 137

8.1 Biological Framework
Bioinformatics is a discipline that binds biology and computer science and
deals with tasks such as the acquisition, storage, analysis and diffusion
of biological data. The data that this field deals include very often DNA
and amino acid sequences. Bioinformatics uses advanced computational
approaches to answer a wide variety of problems in molecular biology.
Traditionally, the alignment of sequences, gene finding, the establishment
of evolutionary relationships or the prediction of three-​dimensional forms
of sequences were the main problems addressed by this field (Mount, 2004).
Recently, the new trends that have appeared include the functional predic-
tion of proteins.
Proteins are complex chemical macromolecules which play a funda-
mental role in life within organisms (O’Connor et al, 2010). They can be seen,
according to their primary structure, as long chains of amino acids connected

DOI: 10.1201/9781003104858-8 131


132 Machine Learning Algorithms and Applications in Engineering

to each other. Amino acids are small organic molecules that are linked by a
peptide bond. These linear chains fold in space forming a three-​dimensional
structure. The way of folding is dependent on the amino acids that compose
it1 and determines the properties of the protein.
The importance of proteins lies in the many functions they perform in
organisms. For example, at the structural level, they make up the majority
of cell material; at the regulatory level, enzymes are proteins; at the immune
level, antibodies are glycoproteins. They constitute more than 50 percent of
the dry weight of cells (Liu, 2020), and it can be generally stated that they are
involved in regulating all the processes that take place in living beings, and
in the vast majority of cases, their molecular function is determined by the
relationship they establish with other proteins (Braun and Gingras, 2012) and
with other molecules in the environment.
Due to how determinant it is to know the function of the different proteins
in an organism, many scientific projects are currently focused on trying to elu-
cidate their behaviour, regulation and possible activity from a biochemical,
biological, biomedical or bioinformatics angle. Hundreds of thousands of art-
icles are published every year describing the activity of proteins in different
situations because it is a crucial factor in analysing cellular mechanisms, iden-
tifying the functional changes that lead to possible problems at a systemic
level and discovering new tools for the prevention, diagnosis and treatment
of diseases. In other words, knowing the function of proteins is synonymous
with understanding life at a molecular level, which has a series of relevant
implications for the pharmaceutical and biomedical industry.
In late December 2019, the World Health Organization (WHO) was noti-
fied of several cases of pneumonia of unknown aetiology, including severe
cases. Shortly after, the new coronavirus SARS-​CoV-​2 (Zhu et al., 2020) was
identified as the causative agent. International health authorities recorded
a rapid spread of the virus, calling the outbreak of SARS-​CoV-​2 infections a
global pandemic and declaring it a Public Health Emergency of International
Importance. Given the seriousness, numerous costly research projects began
to be developed around the pathogen. To date, after genome sequencing and
analysis, several genes of the coronavirus that causes COVID-​19 have been
discovered, which encode the synthesis of 39 proteins (Gordon et al., 2020)
for the time being. In some cases, it has been possible to identify the functions
they play. However, in others, they are still a mystery.
When discussing the function of proteins, it should be borne in mind that
this is a concept that can have different meanings depending on the different
contexts and/​or biological levels. Thus, the biochemical/​molecular level,
the cellular level and the phenotypic level can be considered. Therefore, the
function of a protein can be categorized into three main groups:

• Molecular function
• Biological process
• Cellular component
Accurate Automatic Functional Recognition of Proteins 133

Nevertheless, the description of the functions of proteins tends to be


structured in ontologies, with the purpose of facilitating and standardizing
the use and exchange of concepts and knowledge, especially when faced
with the problem of massively processing that information. An ontology is a
formal definition of types, properties and relationships between entities that
exist in a particular domain. In the case of proteins, the activity is described
by using hierarchical structures such as directed acyclic graphs or trees. In
all these representations, the nodes or terms describe particular activities,
and the links represent relationships between the terms. There are different
ontology projects that describe the functions of proteins. However, the most
popular and widespread is the one developed by the Gene Ontology (GO)
Consortium (Gene Ontology Consortium, 2019).
The function of proteins may also vary in time and space. Numerous cases
demonstrate this. For instance, alternative splicing (Zheng, 2016) is a well-​
studied and well-​known mechanism in which, after transcription, different
exons of the same genome sequence bind in different ways to produce dis-
tinct forms of proteins. Similarly, after translation, proteins may be split into
functional fragments or the chemical modification of various subgroups
(residues) of amino acids can occur, thus changing their properties and, as a
consequence, modifying the overall function of the protein. In total, more than
four hundred post-​translational and post-​translational modification events
are known, resulting in a very wide range of possible functional forms of a
protein, all from the same sequence of amino acids. Another widely known
case is that of ‘moonlighting’ or multifunctional proteins (Jeffery, 2018),
proteins that can perform two completely different functions. Furthermore,
as mentioned above, on many occasions, proteins interact in groups to carry
out separate biological processes. All this has practical implications for the
identification and determination of protein functions, which means that this
problem remains a significant challenge in the scientific community today.

8.2 Identifying the Protein Functions


Functional annotation of proteins is the set of tasks involved in the process of
identifying and assigning labels to the function of known proteins (or groups
of them) in organisms. Traditionally, functional annotation was performed
through in vitro and in vivo experiments, being recorded in databases of bio-
logical sequences through a curation process based on the existing literature.
However, these laboratory experiments and manual curation tasks, within
a biological context of such complexity as described, were excessively hard
in terms of difficulty, inherent costs and time. Advances in technology have
led to significant developments in the functional annotation of proteins, but
at the same time to considerable growth in the size of biological databases.
134 Machine Learning Algorithms and Applications in Engineering

Next-​generation sequencing (NGS) (Slatko et al., 2018) has boosted a spec-


tacular increase in protein sequence databases as the number of genome and
metagenome sequencing projects for many species has grown. It reached a
point where the experimental definition of protein function could not handle
the vast amount of data of sequences available. In other words, as the produc-
tion of new collections of sequences increases exponentially, the effort of the
teams carrying out in vitro or in vivo experimentation, and of the curators,
has not been able in the same way, so that there is a rising number of protein
sequences without their supervised or curated annotation being possible. It
is, therefore, necessary to develop computational tools that take charge of
this work with the greatest possible accuracy. From that moment on, the
panorama invited the development of new functional annotation computa-
tional techniques.

8.3 Automatic Functional Annotation of Proteins


The simplest and most widely used method to perform the task of automat-
ically determining protein functions is based on homology searches. This is
achieved through programs that search for similarities between sequences.
For example, Blast2GO (Götz et al., 2008) is one of the most widely used tools
of this type, making use of the BLAST algorithm (Camacho et al., 2009) to
try to assign to not annotated proteins the function(s) of their homologues,
which have been previously inferred and the function is known. However,
this analysis depends directly on the databases involved and the annotation
of the sequences in them. It is therefore relevant to consider the propagation
of the error, that is, if these databases contain errors, the automatic propaga-
tion of the errors must be expected, or if the database used is too small, the
resulting annotations may be unsatisfactory.
UniProt (UniProt Consortium, 2018) is the consortium that maintains a
centralised database that gathers all the functional information on proteins,
with precise, consistent and rich annotation. This repository, UniProt
Knowledgebase (UniProtKB), is a protein sequence bank consisting of two
sections: UniProtKB/​Swiss-​Prot, which currently contains 562,253 manually
annotated and reviewed entries, and UniProtKB/​TrEMBL, which contains
180,690,447 entries annotated by automatic, unsupervised, computer
techniques and which therefore await full manual annotation.
However, in the situation of an unknown protein that does not have sig-
nificant similarity to any known protein, computational approaches can be
used to predict protein function. These are techniques that use only sequence
or structure information to infer properties that are common to proteins of
the same function. These methods assume that proteins of the same function
adapt similarly to the same conditions. They use information such as the
Accurate Automatic Functional Recognition of Proteins 135

underlying amino acid chain itself (Kulmanov et al, 2018), three-​dimensional


structure (Yang et al, 2015), binding properties, structural flexibility (Peled
et al., 2016), protein–​protein interaction networks (Kulmanov et al, 2018) or
other molecular and functional factors (Rivarola). Although not directly, these
methods may also be partially dependent on protein databases with a known
function. Such methodologies rely on supervised machine learning models
and can find significant correlations between characteristics and functions.
From a computational perspective, the prediction of protein function can be
considered as a multi-​label classification problem (Tsoumakas and Katakis,
2007). In the traditional classification machine learning models, each instance
is associated with only one class. However, multi-​label classification models
deal with an environment in which each instance can be associated with more
than one class at the same time, which makes multi-​label problems more
complex and difficult than their single-​class equivalents, whether binary or
multiple.
Formally, a multi-​label problem can be defined as follows (Spolaôr et al.,
2013): let D be a data set composed of N examples En=​(xn,Yn), n=​1…N. Each
example En is associated with a feature vector xn=​(xn1, xn2...,xnM) described by
M features Xm, m=​1…M, and a subset of labels Yn ⊆ L, where L=​{y1, y2,...,yq} is
the set of q labels. In this scenario, the multi-​label classification task consists
in generating a classifier H, which, given an unseen instance E=​ (x,?), is cap-
able of accurately predicting its subset of labels Y, i.e., H(E)→Y.
The multi-​label problems are more complex and difficult than their single-​
class equivalents, whether binary or multiple. Furthermore, the enormous
scale of the problem highly complicates any data-​driven approach that aims
to address it. Table 8.1 shows the features of the data sets employed for one of
the highest performance methods (Kulmanov et al., 2018) currently available.
It is clear what the true magnitude of the problem is in terms of its dimension.
At the same time, it should be considered the scenario that arises when
dealing with the integration of information from different and heterogeneous
sources, in which it is necessary to deal with various biological data (sequence,
structure, interactions, etc.). Moreover, a good volume of data is not yet avail-
able for most species. That makes the learning process of models with the
sequences of one species complicated to apply when it comes to making
inferences about others. The limitation of performing a manual validation of

TABLE 8.1
Number of Sequences (patterns) with Experimental Annotations in Data Sets Grouped
by Sub-​ontologies

MFO BPO CCO All

Training size 36,110 53,500 50,596 66,841


Testing size 1137 2392 1265 3328
Number of labels 677 3992 551 5220
136 Machine Learning Algorithms and Applications in Engineering

functional assignments has traditionally led to focusing most of the available


resources on species of very high relevance (human, 20,365; common mouse,
17,038; Arabidopsis thaliana, 15,952 or S. cerevisiae, 6721),2 which has meant
that there is a significant gap in the availability of revised functional data
for species considered less essential and/​or urgent. Finally, there is not yet
a clear procedure on how the performance of annotation methods should be
evaluated, since it is necessary to develop good similarity functions between
pairs of subgroups in the ontologies.
Functional protein prediction algorithms, both those based on homologies
and those based on machine learning, are quite expensive from a computa-
tional point of view, both in their development and, especially, in their appli-
cation. Although the increase in computing power available today, whether
CPU-​based or increasingly GPU-​ based, makes it easier to compare new
algorithms and to apply them to larger molecules and molecular assemblies.
In many cases, the number of sequences in the training set is often limited
by the lack of computing resources, which affects the performance of the
methods (Rifaioglu et al., 2019).
After an analysis of the results of the predictions of current methods
(Figure 8.1) (Radivojac et al., 2013), it can be stated that there is still margin
for significant improvement in this field. The critical assessment of functional
annotation (CAFA) is an initiative intended for the large-​scale evaluation of
automatic methods for the functional prediction of proteomes. It is
currently in its fourth edition (2019–​2020 CAFA4), although the first edition
was organised between 2010 and 2011. In the challenge they formulate, they
make available to the community a database of protein sequences whose
functions are not yet empirically known. This blind data set can be addressed
by any computational model to try to predict which functionalities each of
the sequences has. Each participant can use the sources of information and
procedures they consider appropriate to complete this task. After the closing
of the call, an article with the precision of each participating method is made
public. The results of CAFA3 (2016–​2017) and CAFA 3.14–​IP (2017–​2018)

FIGURE 8.1
Overall performance evaluation at two levels.
Accurate Automatic Functional Recognition of Proteins 137

reveal that homology-​based methods are outperformed by those that are


not, but that, nevertheless, the measures of accuracy of the best participating
models indicate a situation that is far from being able to consider the func-
tional annotation of proteins as a closed problem.

8.4 Challenges
In the light of the aforementioned, after an exhaustive review of the specific lit-
erature on the problem of protein function prediction, a number of challenges
have been identified that could be of great interest to scientists involved in
research projects on this topic. First, to propose new assessment metrics for
the problem of automatic functional annotation. Within this point, models
have traditionally been evaluated with metrics derived from precision and
recall. These metrics are often biased towards the majority classes and do not
consider the class hierarchy present in the problem. It would be interesting
to propose new evaluation metrics for supervised classifiers implemented
to solve automatic protein function prediction problems. Within this area,
models have traditionally been evaluated with metrics derived from pre-
cision and recall. Such metrics (e.g., F-​score) tend to have a bias towards
majority classes (as does the correct classification rate [CCR]). It also assumes
that the distributions associated with the class label distributions and the
predicted distributions are equal. In short, the F-​score is not a suitable metric
to measure the goodness of fit of a classifier in a typical bioinformatics scen-
ario with multiple classes and imbalance problems between them. Moreover,
this problem presents hierarchical classes so that classification errors should
not all have the same weight.
Second, it is considered necessary to develop scalable machine learning
models for the problem under study. Functional protein annotation is a classi-
fication problem composed of more than 100k instances, more than 5k classes
and of a multi-​label nature. State-​of-​the-​art machine learning algorithms have
serious computational difficulties in addressing problems of this typology
due to these models scale up with the sample size of the problem. In order
to overcome these serious computational difficulties to deal with problems
with so many instances and classes, it would be appealing to test lightweight
alternatives (if compared to classical algorithms) and propose models whose
computational complexity is adjusted to the needs of the problem. Moreover,
these models will have to be designed taking into account that, on the one
hand, this is a multi-​label problem, and that, on the other hand, they must rely
on different and diverse sources of information, which makes an ensemble
perspective ideal.
Finally, in line with the above challenge, it should again be noted that the
data sets found in the protein function prediction problem contains more than
138 Machine Learning Algorithms and Applications in Engineering

4,000 attributes. Measures of inter-​pattern similarity in machine learning are


usually based on the Euclidean metric. This metric has the problem of con-
centrating distances in high-​dimensionality spaces and motivated by this
drawback, new inter-​pattern similarity non-​Euclidean-​based metrics need to
be proposed for the previously exposed scenario.

Notes
1 The way of folding does not only depend on the sequence of amino acids. For
example, prions are proteins that fold in a different way than native protein but do
not change their amino acid sequence. There are proteins in the cell that are respon-
sible for the correct folding of the generated new proteins. The physicochemical
environment may also affect the final result of the folding.
2 Number of entries in UniProtKB/​Swiss-​prot.

References
Braun, P. and Gingras, A.C. (2012). History of protein–​protein interactions: From egg
white to complex networks. Proteomics, 12(10), 1478-​1498.
Camacho, C., Coulouris, G., Avagyan, V., Ma, N., Papadopoulos, J., Bealer, K. and
Madden, T. L. (2009). BLAST+​: architecture and applications. BMC Bioinformatics,
10(1), 421.
Gene Ontology Consortium (2019). The gene ontology resource: 20 years and still
GOing strong. Nucleic Acids Research, 47 (D1), D330–​D338.
Gordon, D.E., Jang, G.M., Bouhaddou, M. et al. (2020). A SARS-​CoV-​2 protein inter-
action map reveals targets for drug repurposing. Nature 583, 459–​468.
Götz, S., García-​Gómez, J.M., Terol, J., Williams, T.D., Nagaraj, S.H., Nueda, M.J., ...
& Conesa, A. (2008). High-​throughput functional annotation and data mining
with the Blast2GO suite. Nucleic Acids Research, 36(10), 3420–​3435.
Jeffery, C.J. (2018). Protein moonlighting: what is it, and why is it important?
Philosophical Transactions of the Royal Society B: Biological Sciences, 373(1738),
20160523.
Kulmanov, M., Khan, M. A. and Hoehndorf, R. (2018). DeepGO: predicting protein
functions from sequence and interactions using a deep ontology-​aware classi-
fier. Bioinformatics, 34(4), 660–​668.
Liu, S. (2020). Bioprocess Engineering: Kinetics, Sustainability, and Reactor Design. Elsevier.
Mount, D.W. (2004). Sequence and genome analysis. Bioinformatics. Cold Spring
Harbour Laboratory Press: Cold Spring Harbour, 2.
O’Connor C.M., Adams, J.U. and Fairman, J. (2010). Essentials of Cell Biology.
Cambridge, MA: NPG Education, 1, 54.
Accurate Automatic Functional Recognition of Proteins 139

Peled, S., Leiderman, O., Charar, R., Efroni, G., Shav-​Tal, Y. and Ofran, Y. (2016).
Denovo protein function prediction using DNA binding and RNA binding
proteins as a test case. Nature Communications, 7(1), 1–​9.
Radivojac, P., Clark, W.T., Oron, T.R., Schnoes, A. M., Wittkop, T., Sokolov, A. … and
Pandey, G. (2013). A large-​scale evaluation of computational protein function
prediction. Nature Methods, 10(3), 221–​227.
Rifaioglu, A.S., Doğan, T., Martin, M.J., Cetin-​ Atalay, R. and Atalay, V. (2019).
DEEPred: Automated protein function prediction with multi-​task feed-​forward
deep neural networks. Scientific Reports, 9(1), 1–​16.
Slatko, B.E., Gardner, A.F. and Ausubel, F.M. (2018). Overview of next-​generation
sequencing technologies. Current Protocols in Molecular Biology, 122(1), e59.
Spolaôr, N., Cherman, E.A., Monard, M.C. and Lee, H.D. (2013). A comparison of
multi-​label feature selection methods using the problem transformation
approach. Electronic Notes in Theoretical Computer Science, 292, 135–​151.
Tsoumakas, G. and Katakis, I. (2007). Multi-​ label classification: An overview.
International Journal of Data Warehousing and Mining (IJDWM), 3(3), 1–​13.
UniProt Consortium (2018). UniProt: the universal protein knowledgebase. Nucleic
Acids Research, 46(5), 2699.
Yang, J., Yan, R., Roy, A., Xu, D., Poisson, J. and Zhang, Y. (2015). The I-​TASSER
Suite: protein structure and function prediction. Nature Methods, 12(1), 7.
Zheng, S. (2016). IRAS: High-​throughput identification of novel alternative splicing
regulators. Methods in Enzymology (Vol. 572, pp. 269–​289). Academic Press.
Zhu, N., Zhang, D., Wang, W., Li, X., Yang, B., Song, J., … and Niu, P. (2020). A novel
coronavirus from patients with pneumonia in China, 2019. New England Journal
of Medicine, 382 (8), pp. 727–​733.
9
Taxonomy of Shilling Attack Detection
Techniques in Recommender System

Abhishek Majumder, Keya Chowdhury, and Joy Lal Sarkar


Department of Computer Science and Engineering, Tripura University,
Suryamaninagar, Tripura

CONTENTS
9.1 Introduction................................................................................................. 142
9.1.1 Collaborative Filtering in Recommender Systems.................... 143
9.1.1.1 User-Based Collaborative Filtering (UBCF)................. 143
9.1.1.2 Item-Based Collaborative Filtering (IBCF)................... 144
9.2 Profile Injection Attack Method............................................................... 145
9.3 Shilling Attack Detection........................................................................... 146
9.4 Rating........................................................................................................... 146
9.5 Time Interval............................................................................................... 146
9.6 Classification............................................................................................... 147
9.6.1 Using the Rating Parameter.......................................................... 147
9.6.2 Detection of Attack Profiles........................................................... 147
9.6.2.1 Hilbert–Haung Transform and Support
Vector Machine (HHT-SVM)........................................ 148
9.6.2.2 Re-scale AdaBoost......................................................... 148
9.6.2.3 Variable-Length Partitions with Neighbor
Selection.......................................................................... 148
9.6.2.4 Principal Component Analysis.................................... 149
9.6.2.5 Discrete Wavelet Transform and Support
Vector Machine (DWTSVM)......................................... 149
9.6.2.6 Principal Component Analysis and Perturbation..... 149
9.6.2.7 Semi-supervised Learning Using the Shilling
Attack Detection (SEMI-SAD)..................................... 150
9.6.2.8 Support Vector Machine and Target Item
Analysis (SVM-TIA)...................................................... 150
9.6.2.9 Rating Deviation from Mean Agreement................... 150
9.6.2.10 Novel Shilling Attack Detection.................................. 150

DOI: 10.1201/9781003104858-9 141


newgenprepdf

142 Machine Learning Algorithms and Applications in Engineering

9.6.3 Attack Items Detection.................................................................. 151


9.6.4 Attack Profiles and Attack Items Detection................................ 151
9.6.4.1 Double M Detector (DMD) and Heterogeneous
Information Network (HIN) Method............................ 151
9.6.5 Using the Rating and Time Interval Parameter......................... 151
9.6.6 Attack Profiles Detection............................................................... 151
9.6.6.1 Detection Based on the Statistical Model..................... 152
9.6.6.2 Unsupervised Technique Using the Hidden
Markov Model and the Hierarchical Clustering......... 152
9.6.6.3 Attack Detection in Time Series..................................... 152
9.6.7 Attack Items Detection.................................................................. 153
9.6.7.1 Dynamic Time Interval Segmentation Technique....... 153
9.6.7.2 Unsupervised Retrieval of Attack Profiles................... 153
9.6.7.3 Data Tracking.................................................................... 153
9.6.8 Attack Profiles and Attack Items Detection................................ 153
9.6.9 Advantages and Disadvantages................................................... 153
9.7 Conclusion................................................................................................... 157
9.7.1 Future Direction.............................................................................. 158

9.1 Introduction
Within the last 20 years, recommender systems have come out as one of the
efficient techniques to deal with the knowledge, which is overloaded by sug-
gestive information and it acquires potential interest to the online users. They
are helpful to the businesses manufacturing merchandise as it increases the
selling rate, cross-​sales, and customers’ loyalty. As a result, customers tend to
come back to the sites that best serve their desires. The recommendation strat-
egies are typically categorized into three main prospects: (i) content-​based
recommendation, (ii) collaborative filtering, and (iii) hybrid recommenda-
tion approaches. Recommendation system, particularly the collaborative
filtering (CF)-​based system, is introduced with success to filter out irrelevant
resources (Si and Li 2020, Sarwar et al. 2001). In the recommender systems,
collaborative filtering recommender system (CFRS) is considered as one of
the most well-​liked and productive techniques. CFRS works on the principle
that identical users have identical tastes. However, collaborative filtering
results in the source of strength and vulnerability for recommender systems
due to its open and interactive nature. Generally, a user-​based CA algorithm
makes recommendation by searching out similar user patterns, which are
illustrated by the preferences of numerous totally non-​identical people (Si
and Li 2020). If profiles contain biased information, they may be thought as
real users and eventually lead to biased recommendations. Therefore, rele-
vant data gets buried under a good deal of irrelevant information. Collective
Shilling Attack Detection Techniques in Recommender System 143

filtering is one in all guidance strategies, which encourage clients to pick a


relevant item. It is acclimated for taking care of data overload drawbacks
by assembling incredibly right predictions. The key suspicion of collabora-
tive filtering method is that clients having comparable experiences on past
things are having a tendency to concur on new things (Bilge et al. 2014a). In
any case, these are inclined to profile injection or shilling attacks. As per Lam
and Riedl (2004) and O’Mahony (2004), one of the precise negative impacts
found in recommender systems (RSs) is that the “deceitful producers” are
sometimes used to make a decision for obtaining a more deceitful route. The
primary objective of selecting such a route is to influence the RSs in the pro-
cess of item recommendation considering its rating. Such phenomena can
be easily observed from the promotion technique of Sony Pictures for fresh
released films in June 2001. The same technique can be used by Amazon.com
for selling some non-​rated books. Meanwhile, eBay did the same by pur-
chasing good feedbacks (ratings) from other members. Privacy-​preserving
CF (PPCF) systems have been developed to protect such personal preferences
(Bilge et al. 2013). In this chapter, a brief discussion on profile injection tech-
nique has been presented. The profile injection technique primarily works
on two parameters, namely, rating parameter and rating and time interval
parameter.

9.1.1 Collaborative Filtering in Recommender Systems


Collaborative filtering is one-​in-​all advice technique, which helps users to
choose an applicable product. It is accustomed for handling information
overload downside by manufacturing extremely correct predictions. The key
assumption of collaborative filtering technique is that users having similar
experiences on past things tend to agree on new items (Bilge et al. 2014a).
Collaborative filtering systems utilize terribly sparse p × q user-​item matrix,
which has p users’ preferences regarding q merchandise. The systems provide
recommendations to their users by evaluating alternative similar users’ pref-
erence. Collaborative filtering strategies are undefeated at providing correct
referrals regarding merchandise. They are additionally able to overcome data
overload downside by matching users with right things for them. However,
these are prone to profile injection or shilling attacks. There may exist some
users or corporations who are malicious aiming to control the recommender
systems outcomes for their benefits.
The collaborative filtering strategies are typically categorized into two
main prospects: (i) user-​based collaborative filtering and (ii) item-​based col-
laborative filtering.

9.1.1.1 User-​Based Collaborative Filtering (UBCF)


The user-​based collaborative filtering (UBCF) basically works on the strategy
of recommending items to active users by finding similar users. Figure 9.1
144 Machine Learning Algorithms and Applications in Engineering

FIGURE 9.1
Framework of user-​based collaborative filtering (UBCF).

shows the framework of UBCF. For this technique, it needs two tasks. First,
using a similarity function the k-​nearest neighbors (kNN) to the user are
found. The kNN-​based algorithm is the most popular CF algorithm. Data is
represented as a u × i user-​item matrix. Where an entry (u, i) indicates either
the rating user u gave to item i, if he/​she rated it, or null otherwise. Using
the Pearson correlation given in Equation (9.1), it computes the similarity
between users (Chirita et al. 2005) as

Wij =
∑ k ∈I (R ik
 R −R
−Rl jk

j )( ) (9.1)
( )∑ ( )
2 2
∑ k ∈I

Rik − Rl k ∈I

Rjk − R j

where I represents the set of items. Rik and Rjk are the ratings users i and j gave
 and R
to item k. R  are the average ratings of user i and j, respectively.
l j

Finally, using the kNN formula, the predictions for user i and item b can
be computed as

Pia = Ri +

k
j =1 (
Wij Rja − Rj )
(9.2)

k
j =1
Wij

9.1.1.2 Item-​Based Collaborative Filtering (IBCF)


The item-​based collaborative filtering (IBCF) algorithm is based on the item–​
item similarity instead of focusing on the similar user. Figure 9.2 shows the
framework of IBCF. This algorithm needs two tasks. First, calculate the simi-
larity between items i and j [8]‌given in Equation (9.3). There are different
types of strategies for finding the similarity among the items. Other similarity
calculation techniques are cosine-​based similarity, adjusted cosine similarity,
correlation-​based similarity and 1-​jaccard distance.
Shilling Attack Detection Techniques in Recommender System 145

FIGURE 9.2
Framework of item-​based collaborative filtering (IBCF).


i. j
sim ( i , j ) =   (9.3)
i. j

Then, the prediction for an item is computed by using weighted sum. Here,
Rub is a weighted average of user’s ratings. Therefore, the prediction score is
calculated using the following formula given as

∑ all similar items , a


Rua sim (i , a )
Pui = (9.4)
∑ all similar items , a
sim (i , a )

9.2 Profile Injection Attack Method


CFRSs are broadly utilized in e-​ commerce markets, during which an
opponent might try to misguide the recommender framework. Shilling
profiles (SPs) are injected by the attacker to the target items just to expand
or reduce the recommending frequency (Lam et al. 2004). The shilling profile
is also employed as a training sample to work out the recommendations for
genuine clients. Therefore, the target items are nominated with additional
frequency or else less frequencies (Gunes et al. 2014). Table 9.1 displays a gen-
eral shilling profile in a shilling attack.
To eliminate detection, SPs could also be visible as legitimate ones. A shilling
profile (SP) mainly consists of four elements: (i) the items which are selected
(iA) and its rating distribution δ(iA), (ii) filler items (iD) and its rating distri-
bution σ(iD), (iii) the items which are unrated (iα) with null ratings, and
(iv) an item which is targeted (iP) and its rating distribution γ(iP) (Bhaumik
et al. 2006, 2007, 2011). iA is specified for certain attack method. iD is chosen
146 Machine Learning Algorithms and Applications in Engineering

TABLE 9.1
General Shilling Profile

Items iA1 ... iAj iD1 ... iDl iα1 ... iαl iP
Ratings δ(iA1) ... δ(iAj) σ(iD1) ... σ(iDl) Null Null Null γ(iP)

randomly and attackers carefully style their rates to camouflage as legitimate


profiles. iα deals with the non-​rated items. iP is the item given the highest or
lowest rating. It also indicates the item which the offender aims to attack. The
strength of shilling attack (SA) is described by the size of the filler and size of
the attack. The ratio of range of the items, which are rated in a profile to the
whole variety of items, represents the size of the filler. The ratio of the quality
of SP and the total number of all profiles denotes the size of the attack. The
size of the filler and size of the attack are usually not large for the price of
attack and to ignore detection.

9.3 Shilling Attack Detection


To reduce the effect of shilling attacks (SAs) in CFRSs, there are two common
techniques (Si and Li 2020). One technique is SA detection and other tech-
nique is before running CF algorithms excludes the attack profiles from the
rating information. Several schemas are proposed to find such SAs. Another
different approach is to develop attack-​resistant CF algorithm. The SA iden-
tification algorithms will be classified as follows: supervised classification
techniques, unsupervised clustering techniques, semi-​supervised techniques,
and other techniques.

9.4 Rating
Rating is a measurement of the quality of something, especially when
compared with other things of the same type.

9.5 Time Interval
A clock breaks the time into the intervals of hours, minutes, and seconds. An
interval is a discrete measurement of time between two things.
Shilling Attack Detection Techniques in Recommender System 147

9.6 Classification
In this section, a classification of shilling attack detection technique has been
presented. Based on the parameter used, the techniques have been classified
into two types: (i) using rating parameter and (ii) using rating and time
interval parameter. The techniques are further classified into three types
based on the output: (i) attack profile detection, (ii) attack item detection and
attack profiles, and (iii) attack item detection. The classification is shown in
Figure 9.3.

9.6.1 Using the Rating Parameter


In this type of technique, the shilling attack is detected using rating
parameter.

9.6.2 Detection of Attack Profiles


This type of technique is used to detect attack profiles. Examples of this type
of techniques are Hilbert–​Haung transform and support vector machine

FIGURE 9.3
Classification.
148 Machine Learning Algorithms and Applications in Engineering

(HHT-​ SVM), re-​


scale AdaBoost, variable-​length partitions with neighbor
selection (VLPNS), principal component analysis (PCA), discrete wavelet
transform and support vector machine (DWTSVM), PCA and perturbation,
semi-​supervised learning using the shilling attack detection (SEMI-​SAD),
support vector machine and target item analysis (SVM-​TIA), rating deviation
from mean agreement (RDMA), and novel shilling attack detection.

9.6.2.1 Hilbert–​Haung Transform and Support Vector Machine


(HHT-​SVM)
The Hilbert–​ Haung transform and support vector machine (HHT-​ SVM)
technique (Fuzhi and Zhou 2014) has been employed to identify the profile
injection attacks by using the combination of support vector machine and
Hilbert–​Haung transform. The HHT is accustomed to transforming the rating
series of the individual user profile for characterizing the profile injection
attack. The SVM is used to detect profile injection attack. The disadvantage
of this method is as soon as new types of attacks occur, the SVM classifier
requires being re-​trained offline. It is necessary for HHT-​SVM to know about
new kind of attacks incrementally, because in real CFRs’ context, attacks are
changing dynamically.

9.6.2.2 Re-​scale AdaBoost
This detection method improves the overall performance with respect to
following two aspects. Initially, it finds out on well-​structured features from
client properties. Then using the statistical properties based on the various
attack models, it makes hard identification circumstances to act simpler.
Later, with reference to the general thought of re-​scale Boosting (RBoosting)
and AdaBoost, it tends to use a variation of AdaBoost, known as the re-​scale
AdaBoost (RAdaBoost) (Yang et al. 2016). The RAdaBoost detection tech-
nique supports the extracted features. This implies that if the parameters are
selected suitably, the RBoosting technique is basically an optimal boosting-​
type calculation. The RAdaBoost may be employed in conjunction with sev-
eral alternative sorts of the machine learning (ML) algorithms for upgrading
the capability of SA recognition. This technique also increases stress on
concerned attacks. On a troublesome classification task, it will clearly improve
the predictive capability.

9.6.2.3 Variable-​Length Partitions with Neighbor Selection


Variable-​length partitions with neighbor selection (VLPNS) algorithm (Wei
and Shen 2016) is designed to reduce the matter of unassured prediction
accuracy of CFRS techniques against the SAs. It relies upon threshold setting.
VLPNS has the particular property of using soft decision procedure that
Shilling Attack Detection Techniques in Recommender System 149

holds suspicious fakers instead of erasing them. It enables the consideration


of mistakes of the traditional user during similarity calculation. Neighbor
choice with variable-​length allotments create segments with entirely unex-
pected sizes by c-​means grouping. It picks neighbors, which are consistent
with each client’s suspicious possibility. The technique diminishes the likeli-
hood of legitimate clients being detected as fakers. In this way, the prediction
accuracy is increased.

9.6.2.4 Principal Component Analysis


Detecting shilling attacks improves the security of a recommender system
(Lam S K et al. 2004). Principal component analysis (PCA) detection (Deng
et al. 2016) is designed to detect shilling profiles in a recommender system
(Mehta B 2007; Mehta B et al. 2007; Mehta B and Nejdl W 2009). PCA can
be defined as a direct dimensionality decrease method, which generally
intends to project data over another space having lower dimensionality as
well as little co-​variance among the dimensions. High co-​variance among
the dimensions is often viewed as excess information and they need to be
removed. So, the repetition of information is decreased in the original data
at substitution space (Jolliffe 2011). PCA is generally used for finding out
shilling profiles. PCA views clients like dimensions of a client-​item matrix.
In this manner, a profile consisting of very little linear combination of weight
has additional probability of being a shilling profile. PCA detection acts well
enough for identifying the shilling profiles, which are created using attack
models like average attack, bandwagon attack, and random attack (Mehta
B 2007).

9.6.2.5 Discrete Wavelet Transform and Support Vector Machine


(DWTSVM)
The detection technique is basically a hybrid approach by merging the discrete
wavelet transform and the support vector machine (DWTSVM) (Karthikeyan
et al. 2016). To categorize fake or legitimate clients, the DWTSVM method
constructs ratings of each client by utilizing the popularity and novelty
of items rated by the clients. In this method, some nonstationary signals
established by employing the popularity and novelty of items the clients are
rated, which is given as an input to the DWT. The instantaneous frequency,
amplitude and phase of DWT signal are applied to a feature set. Then it is
used by the SVM for classifying the users as fake or genuine users.

9.6.2.6 Principal Component Analysis and Perturbation


The PCA and perturbation (Deng et al. 2016) is an unsupervised shilling
attack (SA) detection technique. PCA is generally applied at the beginning
150 Machine Learning Algorithms and Applications in Engineering

and after application of Gaussian noise in every user profile. Therefore, the
SAs are discovered by merging the results obtained from two PCAs. This
technique achieves higher accuracy in experiments as rating of a shilling pro-
file (SP). The SP contains less deviation from average rating. The effect of
injecting perturbation to SPs ought to be greater than that of the legitimate
profile. The experiment result confirms that this technique outperforms the
initial PCA. Injecting perturbation is useful for shilling attack detection.

9.6.2.7 Semi-​supervised Learning Using the Shilling Attack Detection


(SEMI-​SAD)
The semi-​supervised learning based shilling attack detection (SEMI-​SAD)
technique (Wu et al. 2011) is assisted by shilling attack detection algorithm.
It takes advantage of each kind of information. At first, it trains a small set
of labeled clients as naive Bayes (NB) classifier and then consolidates the
unlabeled clients with expectation maximization (EM) for upgrading the said
NB classifier. The semi-​SAD can more readily distinguish numerous sorts
of shilling attacks compared to others, particularly obfuscated and hybrid
shilling attacks. This technique optimizes naive Bayes in form of the initial
detector and an augmented EM to enhance the detector. This algorithm gains
knowledge from each labeled and unlabeled client profiles. By combining
naive Bayes classifier and unlabeled data, EM is utilized to increase the
weight of the labeled information.

9.6.2.8 Support Vector Machine and Target Item Analysis (SVM-​TIA)


This shilling attack detection technique (Zhou et al. 2016) employs support
vector machine and targets item analysis in the recommender systems. It
works on Border line-​SMOTE strategy, which is utilized to reduce the class
unbalance drawback in taxonomy. Therefore, the fine-​tuning part, which
is used to target items within the potential attack profile set, is analyzed.
This system analyses the ratings of the items to observe the attack profiles.
However, the disadvantage of the technique is that some genuine profiles are
misjudged as an attack profile.

9.6.2.9 Rating Deviation from Mean Agreement


The rating deviation from mean agreement (RDMA) technique (Si and Li
2020; R. Burke et al. 2006) observes attackers though testing of average devi-
ation of the profile for each item. It is weighted by the inverse of the amount
of rating of the item to detect the attack profiles.

9.6.2.10 Novel Shilling Attack Detection


The novel shilling attack detection (Bilge et al. 2014b) method is used
for specific attacks. It uses bisecting k-​means clustering approach, which
Shilling Attack Detection Techniques in Recommender System 151

allocates a binary decision tree (BDT) where attack profiles are gathered
in a leaf node. BDT is constructed by recursively clustering the training
data to locate the fake attack profiles using k-​means clustering algorithm.
The user-​item matrix is divided into two distinct clusters at each level.
Intra-​cluster correlation coefficient is calculated for each internal node.
The process is repeated until there remains at most a predefined number
of users in any leaf node. Then BDT is traversed to detect anomalies with
intra-​cluster correlation coefficient and label the node holding all or most
of the attack profiles.

9.6.3 Attack Items Detection


No technique has been introduced yet, which is designed for attack items
detection.

9.6.4 Attack Profiles and Attack Items Detection


This type of technique is designed to detect attack profiles and attack items.
Examples of this type of techniques are double M detector (DMD) and het-
erogeneous information network (HIN) method.

9.6.4.1 Double M Detector (DMD) and Heterogeneous Information


Network (HIN) Method
The double M detector (DMD) technique detects shilling attack by meta-​path
and matrix factorization. The DMD technique (Zhang et al. 2019) concatenates
the client-​item bipartite network and client–​client relation network. Then it
designs many meta-​paths, which control the stochastic procedure to produce
node sequence and uses the skip-​gram model to get user embedding. To get
the latent factors, matrix factorization is utilized to decompose the client-​item
rating matrix. For representing users’ embeddings of latent relations, numerous
significant meta-​ paths have been designed using heterogeneous informa-
tion network (HIN) with respect to network characteristics, such as coreness,
degree, and hindex. At the end, exploitation embedding and factors are utilized
together to train the identifier to detect the attack profiles and attack items.

9.6.5 Using the Rating and Time Interval Parameter


In this type of technique, the shilling attack is detected using rating and time
interval parameter.

9.6.6 Attack Profiles Detection


This type of technique is used to detect attack profiles. Examples of this type
of techniques are detection based on the statistical model, UD-​HMM, and
attack detection in time series.
152 Machine Learning Algorithms and Applications in Engineering

9.6.6.1 Detection Based on the Statistical Model


The statistical detection algorithm (Wang et al. 2018) first segments the
features of every item with respect to its time-​ordered rating sequence.
Statistical distributions are designed to get the suspicious rating’s relating
time interval. This technique is often drawn as hypothesis test detection
and dynamic time interval segmentation based strategy (Xia et al. 2015).
Based on the central supposition that successful attacks ought to alter the
target item’s statistical properties, this identifier wants to get the irregu-
larity changes of items rating sequence. The hypotheses test is employed
after segmentation of the rating series just to check irregular user groups
among all the groups.

9.6.6.2 Unsupervised Technique Using the Hidden Markov Model


and the Hierarchical Clustering
The shilling attack identification technique relies on the hidden Markov
model (HMM) and hierarchical clustering (Zhang et al. 2018). This tech-
nique measures the difference of user rating behaviors by constructing every
user’s rating item sequence. Then it uses HMM model to get each user’s pref-
erence sequence for calculating each user’s matching degree. It calculates
each item’s entropy to find the users’ suspicious degree for detecting the
attack profiles. The limitation of this technique is that the detection perform-
ance isn’t excellent while detecting the collusive spammer on the sampled
data set.

9.6.6.3 Attack Detection in Time Series


The attack detection in a time series technique (Zhang et al. 2006) uses a
window of size k for building a time series of an item’s rating. It clusters
successive rating of items into disjoint windows. Then it determines the
sample entropy and sample average of the window. It derives a theoretical
optimal size of window for best observation of an attack event if amount of
attack profiles is known. This work begins with the subsequent observation.
It assumes that the attack profiles are infused into the framework in a com-
paratively less amount of time. Most of the shilling attack models share a
common trend in spite of decent variety of the attack amount they induct in
the target items rating distributions (probably for different items). Moreover,
in a large-​scale recommender framework, the ratings of a given item may
have seasonality and/​or pattern over time. A lot of complicated models are
required to include such trends.
Shilling Attack Detection Techniques in Recommender System 153

9.6.7 Attack Items Detection


This type of technique is used to detect attack profiles. Examples of this type
of techniques are dynamic time interval segmentation technique, unsuper-
vised retrieval of attack Profiles (UnRAP), and data tracking.

9.6.7.1 Dynamic Time Interval Segmentation Technique


The dynamic time interval segmentation technique (Xia et al. 2015) is
employed for locating out items attacked by malicious profiles directly.
The entire framework of this technique contains two parts, that is, time
interval segmentation and abnormal interval detection. The time interval
segmentation technique segments the life cycle of every item into many
time intervals dynamically at a checkpoint. The abnormal interval detec-
tion method detects an anomalous item through hypothesis testing. This
detection technique improves the detection performance against target
shifting attack.

9.6.7.2 Unsupervised Retrieval of Attack Profiles


This unsupervised retrieval of attack profiles (UnRAP) detection technique
(Si and Li 2020) is an unsupervised method. This technique detects shilling
attack by analyzing user profiles rating deviation on target time. This tech-
nique can solely detect the attack user profiles of individual time.

9.6.7.3 Data Tracking
This data tracking detection technique (Qi et al. 2018) is used in huge infor-
mation atmosphere. This technique supports new information feature. The
technique uses extended Kalman filter, which quickly tracks and accurately
predicts the rating status of the item supported two new detection attributes,
short-​term average change activity (SACA) and short-​term variance change
activity (SVCA). The detector then detects the abnormal item by comparing
predicted ratings and actual ones.

9.6.8 Attack Profiles and Attack Items Detection


No technique has been introduced yet, which is based on detection of attack
profiles and attack items as output.

9.6.9 Advantages and Disadvantages


In this section, the advantages and disadvantages of the shilling attack detec-
tion techniques have been presented in Table 9.2.
154 Machine Learning Algorithms and Applications in Engineering

TABLE 9.2
Advantages and Disadvantages

The shilling attack


detection technique Advantages Disadvantages

1. Novel shilling attack 1. This method gives promising 1. This method detects
detection ( Bilge et al. results for almost all filler only specific three
2014b). size and attack size. types of attacks namely
bandwagon, segment, and
average attack.
2. When all the profiles are
genuine in the system, this
technique misjudges the
genuine profiles as malicious
profiles.
2. Hilbert–​Haung 1. This method decomposes 1. In this method, as soon as
Transform and each rating series. It also new types of attacks occur,
support vector extracts Hilbert spectrum-​ the SVM classifier requires to
machine (HHT-​SVM) based features for be re-​trained offline.
( Fuzhi et al. 2014). characterizing the profile
injection attacks.
2. Here SVM distinguishes the
genuine users’ profile and
attacker profile.
3. Re-​scale AdaBoost 1. This method extracts 1. The detection rate of this
(Yang et al. 2016). well-​designed features method is low for the attacks
from user profiles. For that have small size of attack
improving detection and filler size.
performance, these profiles 2. Re-​scale AdaBoost cannot
are established based on effectively detect Power User
the various attack models’ Attack-​Aggregate Similarity
statistical properties. (PUA-​AS), Power User
2. It makes hard detection Attack-​Number of Ratings
scenarios easier to perform. (PUA-​NR) and Power User
Attack-​In Degree (PUA-​ID)
attacks.
3. Generic features and type-​
specific features which are
present in this technique
as extractive features are
not enough to depict their
material characteristics.
4. Rating Deviation 1. This algorithm successfully 1. It is unable to detect
from Mean detects attack profiles which segment attack and love/​hate
Agreement (RDMA) are random, average and attack.
(Si and Li 2020; Burke bandwagon.
et al. 2006).
Shilling Attack Detection Techniques in Recommender System 155

TABLE 9.2  (Continued)


Advantages and Disadvantages

The shilling attack


detection technique Advantages Disadvantages

5. Variable-​length 1. Instead of detecting the 1. VLPNS has a higher time


partitions with suspicious fakers, this complexity.
neighbor selection technique marks them and 2. This algorithm does not
(VLPNS) (Wei and thereby reduces false-​positive improve the false negative.
Shen 2016) rate. This process is carried
out in such a way that the
misclassified normal users
will be able to still contribute
to the similarity calculation.
As a result of this, the
probability of mistaking
normal users as fakers
reduces.
2. The prediction accuracy of
this method is high.
6. Principal component 1. PCA performs very efficiently 1. The applicability of PCA
analysis (PCA) (Deng against standard attacks. is limited by certain
et al. 2016). 2. This detection method assumptions made in its
performs very well to derivation.
identify shilling profiles in 2. Another limitation for this
bandwagon attack, random method is the mean-​removal
attack, and average attack process before constructing the
models. co-​variance matrix for PCA.
7. Discrete wavelet 1. This method forms rating 1. It applies Nyquit’s rule. As
transform and series of individual users. The a result, half of the signal
support vector rating series is formed with frequencies are ignored while
machine (DWTSVM) reference to the popularity and sending through the filters.
(Karthikeyan et al. novelty of items. This classifies 2. DWT is computationally
2016) the users as fake or genuine intensive. It is less efficient
users based on time, so that and natural.
the detection rate is easy. 3. The dimensions of SVMs may
2. For getting the feature set, be very high.
discrete wavelet transform is
applied on the rating series.
It is further used by support
vector machine for the
purpose of classification.
8. Principal component 1. By utilizing PCA detection 1. In this method the
analysis and technique, the attack and applicability of PCA
perturbation (Deng legitimate profiles can be is limited by certain
et al. 2016). separated out. assumptions made in its
2. This method is effective, derivation.
simple, and can be easily 2. When perturbation
implemented. accuracy is decremented,
the state values have higher
complexity.
(continued)
156 Machine Learning Algorithms and Applications in Engineering

TABLE 9.2  (Continued)


Advantages and Disadvantages

The shilling attack


detection technique Advantages Disadvantages

9. The semi-​supervised 1. This technique improves 1. In this method classification


learning using the naïve Bayes classifier of any big data is a real
shilling attack by using expectation challenge.
detection (SEMI-​ maximization (EM) to detect 2. Training needs a lot of
SAD) (Wu et al. the user profiles. computation time.
2011). 2. Semi-​SAD is very efficient
and effective.
10. Support vector 1. This method eases the 1. In this method, potential
machine and target class unbalance problem in attack profiles’ set are
item analysis (SVM-​ classification. analyzed, due to which
TIA) (Zhou et al. some genuine profiles
2016). are misjudged as attack
profiles.
11. Double M detector 1. DMD utilizes the 1. This system requires
(DMD) and interconnection between measurement of the
heterogeneous client–​item and client–​client relatedness of different
information network using HIN. It enables the types of objects for
(HIN) method shilling attack identification recommendation.
(Zhang et al. 2019). for CFRSs.
2. This method works
effectively in detecting
the rating attacks as well
as paying attention to the
relation attacks.
3. DMD is also effective for
hybrid attacks.
12. Detection based on 1. A statistical model is 1. This model requires some
statistical model idealized form of items rating information related to
(Wang et al. 2018) generator process. rating distribution of the
2. The probability of any event data set.
can be calculated by the 2. The consumption of time
statistical assumption which gets affected as the scale of
constitutes statistical model. the data set increases.
3. Difficulty in handling leads
to time bias and length-​
biased sampling.
13. Unsupervised 1. Since this technique takes 1. This method does not
method using consideration of current state perform efficiently when
the hidden only, it does not require any detecting the collusive
Markov model prior knowledge. spammer on the sampled
and hierarchical 2. It uses HMM, clusters data set.
clustering (UD-​ genuine users and attackers
HMM) (Fuzhi et al. based on ratings.
2018).
Shilling Attack Detection Techniques in Recommender System 157

TABLE 9.2 (Continued)
Advantages and Disadvantages

The shilling attack


detection technique Advantages Disadvantages

14. Attack detection in 1. The time series of these 1. During attack, sometimes
time series (Zhang sample average and sample normal users’ rating patterns
et al 2006). entropy features can expose also change, so it becomes
attack events by giving difficult to identify shilling
reasonable assumptions attackers.
about their duration.
2. An optimal window size will
be derived theoretically, only
if the attack profile number
is known. This is done to
best identify the rating
distribution changes, which
is caused by attacks.
15. Dynamic 1. This technique divides the 1. In this method, there
time interval life cycle of each item into exists an issue of
segmentation several time intervals. This improving the detection
technique (Xia et al. division is done dynamically performance against target
2015). at checkpoints. shifting attack.
2. It detects anomalous item 2. It is not able to minimize the
through hypothesis test. impact of attacks effectively,
after identifying suspicious
intervals of each item.
16. UnRap (Si and Li 1. This method is used to detect 1. It detects only the attack user
2020). shilling profiles by analyzing profiles for individual times.
user profile’s rating deviation
on target item.
17. Data tracking (Qi 1. Big Data processing is 1. Data tracking is not
et al. 2018) adapted in this technique. applicable to handle large
2. Detection efficiency of the data for distributed system.
data tracking method is high.

9.7 Conclusion
Recommender system (RS) is an application that helps users to select relevant
products from the internet. To reduce the damage of a shilling attack and
maintain good quality of recommendation, the security of RSs is a significant
issue. However, fake user profiles created by SA methods can be accurately
detected by recent SA detection techniques. The ratting pattern of these SA
techniques is different from real users. Being one of the most efficient ways
for handling the problem of information overload, CFRSs are very much
158 Machine Learning Algorithms and Applications in Engineering

vulnerable to numerous shilling attacks due to insertion of variety of mali-


cious user profiles in the system, which effects the user recommendations.
In this chapter, at first the profile injection attack technique, collaborative
filtering in recommender systems and shilling attack detection scheme have
been discussed. Second, the shilling attack detection techniques have been
classified based on the two parameters, namely, rating parameter and rating
and time interval parameter on which they work. The shilling attack detec-
tion techniques have also been classified based on their output algorithms like
attack profiles detection, attack items detection and attack profiles, and attack
items detection. Then each of the shilling attack detection techniques have
been discussed in brief. The techniques are analyzed and their advantages
and disadvantages compared.

9.7.1 Future Direction
The future directions are working on the shortcomings of these shilling
attack detection techniques. Like just in case of dynamic time interval seg-
mentation technique, the detection performance against target shifting attack
requires to be enhanced. A shortcoming of data tracking technique is that it
is not applicable on distributed system to handle giant data. The problem
with HHT-​SVM is that, when new kinds of attacks are conducted, it becomes
difficult to detect attack profiles. SVM classifier also requires being re-​trained
offline. Overcoming these problems are the future directions.

References
Bhaumik R, Burke R, Mobasher B (2007). Effectiveness of crawling attacks against
web-​based recommender systems. Proceedings of the 5th workshop on intelli-
gent techniques for web personalization (ITWP-​07).
Bhaumik R, Mobasher B, Burke R. (2011). A clustering approach to unsupervised
attack detection in collaborative recommender systems. Proceedings of the
International Conference on Data Mining (DMIN). The Steering Committee of
the World Congress in Computer Science, Computer Engineering and Applied
Computing (World Comp), p. 1.
Bhaumik R, Williams C, Mobasher B, Burke R (2006 Jul 16). Securing collaborative
filtering against malicious attacks through anomaly detection. In Proceedings of
the 4th Workshop on Intelligent Techniques for Web Personalization (ITWP’06),
Boston. Vol. 6, p. 10.
Bilge A, Gunes I, Polat, H. (2014a) Robustness analysis of privacy-​preserving model-​
based recommendation schemes. Expert Syst Appl, 41(8): 3671–​3681.
Bilge A, Ozdemir Z, Polat H. (2014b). A novel shilling attack detection method. Proc
Computer Science, 31: 165–​174.
Shilling Attack Detection Techniques in Recommender System 159

Bilge A, Polat H. (2013) A comparison of clustering-​based privacy-​preserving collab-


orative filtering schemes. Appl Soft Comput, 13(5): 2478–​2489.
Burke R, Mobasher B, Williams C. (2006). Classification features for attack detection
in collaborative recommender systems. International Conference on Knowledge
Discovery and Data Mining, pp. 17–​20.
Chirita P-​A, Nejdl W, Zamfir C. (2005). Preventing shilling attacks in online recom-
mender systems. In Proceedings of the 7th annual ACM international workshop
on Web information and data management, pp. 67–​74.
Deng, Z. J., Zhang, F., & Wang, S. P. (2016, July). Shilling attack detection in collab-
orative filtering recommender system by PCA detection and perturbation.
In Proceedings of International Conference on Wavelet Analysis and Pattern
Recognition (ICWAPR), pp. 213–​218.
Fuzhi Z, Zening Z, Peng Z (2018). UD-​HMM: An unsupervised method for shilling
attack detection based on hidden Markov model and hierarchical clustering.
Knowledge-​Based Systems, 148: 146–​166.
Fuzhi Z, Zhou Q. (2014). HHT–​SVM: An online method for detecting profile injec-
tion attacks in collaborative recommender systems. Knowledge-​Based Systems 65:
96–​105.
Gunes I, Kaleli C, Bilge A, et al. (2014). Shilling attacks against recommender systems: a
comprehensive survey. Artificial Intelligence Review, 42(4): 767–​799.
Jolliffe I. (2011). Principal Component Analysis. Springer, Berlin, pp. 1094–​1096.
Karthikeyan P, Selvi ST, Neeraja G, Deepika R, Vincent A, Abinaya V. (2017, January).
Prevention of shilling attack in recommender systems using discrete wavelet
transform and support vector machine. In 2016 Eighth International Conference
on Advanced Computing (ICoAC) IEEE, pp. 99–​104.
Lam SK, Riedl J (2004). Shilling recommender systems for fun and profit. Proceedings
of the 13th international conference on World Wide Web. ACM: pp. 393–​402.
Mehta B (2007). Unsupervised shilling detection for collaborative filtering. AAAI. pp.
1402–​1407.
Mehta B, Hofmann T, Fankhauser P. (2007) Lies and propaganda. Detecting spam
users in collaborative filtering[C]‌. Proceedings of the 12th international confer-
ence on intelligent user interfaces. ACM: pp. 14–​21.
Mehta B, Nejdl W (2009). Unsupervised strategies for shilling detection and robust
collaborative filtering [J]‌. User Modeling and User-​Adapted Interaction, 19(1–​
2): 65–​97.
O’Mahony, Michael P. (2004). Towards robust and efficient automated collaborative
filtering. PhD diss., University College Dublin.
Qi L, Huang H, Wang P and Wang R. (2018, August). Shilling Attack Detection Based
on Data Tracking. In 2018 13th International Conference on Computer Science
and Education (ICCSE), IEEE, pp. 1–​4.
Sarwar, B., Karypis, G., Konstan, J., & Riedl, J. (2001, April). Item-​based collaborative
filtering recommendation algorithms. In Proceedings of the 10th international
conference on World Wide Web: pp. 285–​295.
Si, M., & Li, Q. (2020). Shilling attacks against collaborative recommender systems: a
review. Artificial Intelligence Review, 53(1), 291–​319.
Wei R, Shen H. (2016). An Improved Collaborative Filtering Recommendation
Algorithm against Shilling Attacks. 17th International Conference on Parallel
and Distributed Computing, Applications and Technologies, pp. 330–​335.
160 Machine Learning Algorithms and Applications in Engineering

Wang Y et al. (2018). A comparative study on shilling detection methods for trust-
worthy recommendations. Journal of Systems Science and Systems Engineering
27(4): 458–​478.
Wu Z, Cao J, Mao B, Wang Y. (2011, October). Semi-​SAD: applying semi-​supervised
learning to shilling attack detection. In Proceedings of the fifth ACM conference
on Recommender systems. ACM. pp. 289–​292.
Xia H., Fang B, Gao M, Ma H, Tang Y, Wen J. (2015). A novel item anomaly detec-
tion approach against shilling attacks in collaborative recommendation systems
using the dynamic time interval segmentation technique. Information Sciences
306: 150–​165.
Xin, Z, Hong X, Yuqi S (2019). Meta-​path and matrix factorization based shilling
detection for collaborate filtering. International Conference on Collaborative
Computing: Networking, Applications and Worksharing. Springer, Cham,
pp. 3–​16.
Yang, Z., et al. (2016). Re-​scale AdaBoost for attack detection in collaborative filtering
recommender systems. Knowledge-​Based Systems 100: 74–​88.
Zhang, S, Chakrabarti A, Ford J, Makedon F (2006). Attack detection in time series for
recommender systems. In Proceedings of the 12th ACM SIGKDD international
conference on Knowledge discovery and data mining, ACM. pp. 809–​814.
Zhou W et al. (2016). SVM-​TIA a shilling attack detection method based on SVM and
target item analysis in recommender systems. Neurocomputing 210: pp. 197–​205.
newgenprepdf

10
Machine Learning Applications in
Real-​World Time Series Problems

Antonio Manuel Durán-​Rosal1,* and David Guijo-​Rubio2


1
Department of Quantitative Methods, Universidad Loyola Andalucía, Córdoba, Spain
Department of Computer Science and Numerical Analysis, University of Córdoba,
2

Córdoba, Spain
[email protected]
*

CONTENTS
10.1 Introduction................................................................................................. 161
10.2 Real-World Applications of TSDM Using ML algorithms.................... 164
10.2.1 Massive Missing Data Reconstruction in Wave
Height Time Series......................................................................... 164
10.2.2 Detection and Prediction of Tipping Points............................... 165
10.2.3 Prediction of Fog Formation......................................................... 167
10.2.4 Prediction of Other Convective Situations................................. 169
10.3 Summary and Other Related Works........................................................ 170

10.1 Introduction
This first section introduces the topic presented and the related state-​of-​the-​
art developments. Time series data mining (TSDM) mainly consists of the
following tasks: anomaly detection (Blázquez-​García et al., 2020), classifica-
tion (Ismail-​Fawaz et al., 2019), analysis and preprocessing (Hamilton, 1994),
segmentation (Keogh et al., 2004), clustering (Liao, 2005) and prediction
(Weigend, 2018). More concretely, this chapter is focused on the applications
of time series preprocessing, segmentation and prediction to real-​ world
problems.
Time series analysis and preprocessing are considered previous steps
for other TSDM tasks. One of the most important tasks is the imputation
of missing values in time series, which is essential for the application of

DOI: 10.1201/9781003104858-10 161


162 Machine Learning Algorithms and Applications in Engineering

succeeding methods successfully. It is very common that real-​time series,


especially those collected by sensors, buoys or real physics systems, can
have gaps of information given that they may stop working due to different
reasons. Within this context, many applications have been proposed in recent
years. Typically, the imputation of missing values has been done by inter-
polating the observational feature space, without considering any hidden
dynamics. Zhou and Huang (2018) proposed a model that would capture the
latent complex with a novel iterative imputing network and demonstrated
its performance by its application to a meteorological benchmark data set
outperforming the state-​of-​the-​art methods. More recently and following
the idea of capturing the dynamics from time series, Belda et al. (2020) have
developed a software called DATimeS, which uses machine learning fitting
algorithms such as Gaussian process regression (GPR). It has resulted in
powerful software for the reconstruction of image vegetation time series,
which tend to be discontinuous because of cloud cover.
Time series segmentation consists in the division of a time series into sev-
eral consecutive subsequences, trying to achieve different objectives. More
formally, they are defined as follows: given a time series Y = { yi } , for the time
indexes i = 1 N , the procedure attempts to separate the values of Y into
a set of consecutive l segments. In this way, the time indexes are split into
{
s1 = { y1 ,  , yt1 } , s2 = { yt1 ,  , yt 2 } ,  , sl = ytl−1 ,  , y N , where the ascending }
ordered cut points are expressed as t1 < t2 <  < tl −1 . Considering the

FIGURE 10.1
Segmentation procedure for a time series of length N =​20 and four segments.
s1 = { y1 , y 2 , y 3 , y 4 , y 5 , y 6 } , s2 = { y 6 , y7 , y 8 , y 9 } , s3 = { y 9 , y10 , y11 , y12 , y13 , y14 , y15 , y16 , y17 } and

s4 = { y17 , y18 , y19 , y 20 } , resulting from cut points t1 = 6 , t2 = 9 and t3 = 17 .


ML Applications in Real-World Time Series Problems 163

transition from one segment to the next, each cut point belongs to two
segments, in such a way that they belong to the previous and to the next
segment. Figure 10.1 graphically represents a segmentation procedure.
As abovementioned, segmentation is used to achieve different object-
ives. In the literature, there are two main groups of objectives. On the one
hand, this technique is carried out to discover segment similarities or useful
patterns over time. Methods within this context try to optimize the division
of time series and then group the segments into different clusters. Several
methods have been proposed during the last two decades. Initially, Abonyi
et al. (2003) stated that all points in a time series belonging to the same cluster
are contiguous in time. After that, many authors proposed methods to group
segments instead of points. Tseng et al. (2009) developed an algorithm in
which similarities in wavelet space are considered to guide the segmentation,
resulting in a clustering of segments of different length. Besides, a signifi-
cant clustering of subsequent time series was addressed using two efficient
methods (Rakthanmanon et al., 2012). As can be seen, all these segmenta-
tion procedures require clustering algorithms, given that they aim to make
groups of segments in order to discover useful similarities. In this sense,
several algorithms for time series clustering have been proposed recently
(Guijo-​Rubio et al., 2020a), aiming to obtain time series groups with similar
characteristics or based on segments typologies.
On the other hand, time series segmentation is also applied for approxi-
mating time series; in other words, to reduce the number of time series points.
In this case, the procedure is performed by selecting those points whose
approximation is the most accurate regarding the real-​time series, that is, the
approximation aims to minimize the information loss or the approximation
error. One of the main purposes is to mitigate the difficulty of processing and
memory requirements. A well-​known approach in this context is the use of
piecewise linear approximations, where linear regressions or interpolations
are used for modelling each segment (Keogh et al., 2004). Moreover, Fu
(2011) presented an approach for approximating time series by defining it as
an optimization problem that could be solved by evolutionary algorithms.
Lately, other authors have developed a novel approach based on connected
lines under a predefined maximum error bound (Zhao et al., 2016).
Finally, the prediction of a time series is easily and formally defined as
follows: given a time series Y = { yi } (i = 1 N ) , the prediction consists in the
determination of the value y n + t , T being a future instant of time. This pro-
cedure learns from the known past values in order to generate a model able to
estimate future ones accurately. Traditional statistical models are still widely
used for this task, for example, the COVID-​19 pandemic spread in Saudi
Arabia has been forecasted by AutoRegressive Integrated Moving Average
(ARIMA) models (Alzahrani et al., 2020). Nevertheless, nowadays there has
been an increasing interest in developing novel ML algorithms, such as arti-
ficial neural networks (ANNs), or even more advanced approaches, such as
164 Machine Learning Algorithms and Applications in Engineering

evolutionary ANNs (EANNs) or long short-​term memory (LSTM), which


have been applied to finances (Chandra and Chand, 2016) and to weather
prediction (Karevan and Suyken, 2020), respectively.
The following section presents a set of real-​world applications tackled by
the authors of this chapter.

10.2 Real-​World Applications of TSDM Using ML


algorithms
10.2.1 Massive Missing Data Reconstruction in Wave
Height Time Series
Oceanography buoys are instruments used to measure different wave
properties. Their accuracy and availability are crucial for several essential
operations, like coastal structures maintenance, safe ship navigation or wave
height converters, among others (López et al., 2013). Although there are
many agencies offering support or maintaining the buoys, as is the case of the
National Data Buoy Centre (NDBC) in the United States, many unexpected
events make buoys to break down (e.g. prolonged storms, harmful accidents,
long maintenance periods and so on). In this way, it causes data gaps until
the buoy is completely repaired. Even though some methods can be applied
even with the presence of missing data in the time series, the number of
which need complete data is much higher (Thomson and Emery, 2014). For
this reason, the reconstruction of missing values in time series has attracted a
great interest in this field.
In the last two decades, there has been an exponential increase in the
number of works applying ML techniques to data recovery. One of the first
works in the area was done by Bhattacharya et al. (2003), in which ANNs are
used to retrieve missing values of wave height data in the North Sea. Then,
feedforward multilayer perceptrons (MLPs) and recurrent neural networks
(RNN), trained by conjugate gradient, were presented as accurate methods
for reconstructing the wave height in Turkey (Balas et al., 2004). Besides, a
back-​propagation algorithm for training ANNs was also used in the Atlantic
area with the same purpose of recovering data from buoys (Gunaydin and
Panchang, 2008). More works used ANNs and novel variants during these
years, such as ANNs combined with K-​nearest neighbour algorithm (Zamani
et al., 2008) or ANNs trained by the rough set theory (Setiawan et al., 2008),
among others.
Furthermore, ANNs have also been combined with some other method-
ologies such as the work of Londhe (2008) where ANNs are combined with
genetic programming to estimate missing wave height using neighbours’
stations. In Mahjoobi et al. (2008), a neuro-​ fuzzy inference system was
ML Applications in Real-World Time Series Problems 165

successfully applied over Lake Ontario. And finally, physical methods have
also been combined with ML approaches for wave height estimation (Casa-​
Prat et al., 2014).
As most ML models lack interpretability inherent to these techniques, we
proposed a new method in Durán-​Rosal et al. (2015). In this paper, a product
unit neural network trained by an evolutionary algorithm (EPUNN) was
proposed through a two-​staged procedure. This methodology was applied to
six buoys located at the Gulf of Alaska, United States.
The first stage consisted in performing an initial reconstruction using
transfer functions and neighbour correlation method. On the one hand,
transfer functions are based on the analysis of the correlation and the estima-
tion of a gappy time series by means of a complete one (note that the correl-
ation between them needs to be higher than a predefined threshold). On the
other hand, the neighbour correlation method estimates these missing values
by adding information from the highly correlated buoys.
Once both methods have recovered the missing values, we keep the best
one, that is, the one resulting in the smallest error. After that, the best recovery
for each time series is used as input for the EPUNN of the second stage. In this
sense, the two most correlated inputs (concerning the original time series) are
used to get the definitive reconstructed time series.
The results achieved for the six buoys located at the Gulf of Alaska indicated
that EPUNNs are suitable for this kind of problems, given their accuracy and
their interpretability. Besides, they can be represented as linear models by
applying natural logarithm to the inputs for easing the understanding and
interpretation. As a remark, better reconstructions are achieved for coastal
buoys, given that the availability of values was higher.

10.2.2 Detection and Prediction of Tipping Points


Palaeoclimatology is a popular field of palaeogeography science; it consists
of the study of Earth’s climate characteristics, and concretely, it aims to pro-
vide the best possible description of the climate for a certain period of the
history. Moreover, palaeoclimatology analyses significant climate variations
and the reasons behind them. Some points of no return, thresholds and phase
changes are widespread in nature and often nonlinear. These could have a
high impact on the Earth’s climate being difficult to anticipate (Wassmann et
Lenton, 2012).
Focusing on this problem, several researchers on dynamic systems, such as
climate systems, claim that they present critical transitions. These transitions
are commonly called tipping points (TPs), also known as ‘small things that
can make a big difference’. When a small change in a climate system causes
a strong non-​linear response in its internal dynamics, leading to a change in
its future state, it is a climate TP. In this sense, the main climatic changes on
Earth are determined by the transition points in the temperature time series,
166 Machine Learning Algorithms and Applications in Engineering

or one of their proxies such as the concentration of oxygen isotopes in the


glaciers, which makes them go from one stable state to another.
Early detection of TPs by analysing their causes and what happens before
them has been considered of significant impact since they can affect millions
of lives. Lenton (2011) proposed the differentiation of many types of TPs and
presented some indicators that may help to detect them, like increased auto-
correlation of series values. Besides, Dakos et al. (2012) introduced more par-
ticular techniques related to data processing and indicators. They studied
several methods using simulated ecological data and arrived at principal
conclusions: there is no single best indicator to detect a transition and that all
methods require specific data processing. The previous works have two main
disadvantages: the need of intensive preprocessing of the data and the appli-
cation of a specific treatment for each type of TP. In this regard, we proposed
two works.
In the first work, a genetic algorithm (GA), in combination with an
unsupervised ML technique, was proposed as a segmentation algorithm to
correctly identify the group of segments, where some of them represented TP
transitions (Nikolaou et al., 2015). Based on a random division, the algorithm
optimizes the cut points aiming to improve the quality of the clustering,
according to their statistical similarities (variance, asymmetry, kurtosis,
slope, approximation error and autocorrelation).
This methodology was tested not only using synthetic data sets
generated from recognized dynamic systems but also on two real-​world
paleoclimate oxygen isotope datasets: GISP2 (Greenland Ice Sheet Project
Two) and NGRIP (North Greenland Ice Core Project) δ 18 O time series
with a resolution of 20 years and an average of 5 points to reduce short
fluctuation. The experiments resulted in an effective detection and dis-
covery of similarities and differences in the dynamic characterization of
TPs, called Dangaard–​Oeschger (DO) events in these time series. The algo-
rithm determined that the increase in autocorrelation, variance and mean
squared error can be considered as warning signals for almost all TPs.
Moreover, the proposed approach provided a novel visualization tool for
climate time series analysis.
In the second work, we introduced some improvements and developed
a prediction model derived from the obtained segments (Pérez-​Ortiz et al.,
2019). Once the segmentation stage finished, its evaluation was automatic-
ally done by performing a comparison with an ideal segmentation carried
out by experts in the field and by measuring the algorithm stability. The
algorithm was also supplied with 10 metrics for the clustering validation, in
order to increase the robustness in the search for the best segmentation, thus
resulting in a better performance in terms of TPs detection. For the prediction
phase, the segments obtained by the GA were then transformed into a binary
problem, with the goal of determining whether the next state represented a
DO event.
ML Applications in Real-World Time Series Problems 167

The results achieved confirmed that the improvements made outperformed


those of the previous algorithm detecting all but two TPs in both data sets.
The experiments also demonstrated that the proposed diagnostic-​predictive
ML-​based model, more specifically a decision tree, was feasible and generated
very promising results. This behaviour was produced given that the system’s
prediction-​monitoring period lasts several decades, and that the time series
already contained in its record abrupt climate changes of the same type. The
above-​mentioned aspects open an interesting field for future research.

10.2.3 Prediction of Fog Formation


It is widely known that aviation is one of the means of transport most
affected by adverse weather conditions. There are a lot of phenomena redu-
cing the visibility, such as sandstorms or torrential rains. However, without
any doubt, the most important and most common is fog formation. Fog is a
meteorological phenomenon consisting in the suspension of almost micro-
scopic particles of water droplets in the air, causing the reduction of hori-
zontal visibility on the Earth’s surface. It has been addressed from different
points of view (Román-​Gascón, 2015).
Foggy weather may result in air traffic flow disruptions, flight delays or
even accidents, among others. More specifically, it has a major impact on the
safety and efficiency of airport operations such as taxiing, take-​off and landing
operations (Bergot et al., 2007). Intending to ensure an enough efficiency, air-
port workers need to have the most reliable information, with the aim of
anticipating these phenomena. Some studies have proposed developments
of trustworthy systems to detect and predict fog formation.
Perhaps one of the most widespread methods of predicting fog formation is
the use of numerical weather prediction (NWP) models (Román-​Gascón et al.,
2016). Nevertheless, accurate NWP models are computationally expensive.
In recent years, ML methods have been successfully applied to fog prediction
problems. Fabbian et al. (2007) proposed an ANN to predict fog events on an
18-​hour schedule at Canberra International Airport in Australia. Similarly,
Marzban et al. (2007) used ANNs to combine three sources of information
to forecast ceiling and visibility at 39 airports in the northwestern United
States. In the same year, Bremnes and Michaelides (2007) applied another
ANN approach to fog formation prediction with a time horizon of up to 6 h
in Norway. Then, a Bayesian model was successfully introduced in a fog pre-
diction problem with a time horizon of 12 h in the Pacific Northwest of the
United States (Chmielecky and Raftery, 2011). More recently, Colabone et al.
(2015) proposed an MLP that uses meteorological data to help in the planning
of flight activities of the Academia da Força Aérea (AFA). In Cornejo-​Bueno
et al. (2017), different ML regression techniques have been applied to fog pre-
diction in Spain. These works tackled the problem of fog formation as regres-
sion approaches.
168 Machine Learning Algorithms and Applications in Engineering

Although there are many sensors collecting meteorological data at airports,


such as wind speed and direction, temperature, humidity and pressure,
among others, in order to support the airport operations, the runway visual
range (RVR) is the most important. RVR is a weather variable defined as the
range over which the pilot of an aircraft on the centreline of a runway can see
the markings on the runway surface or the lights that outline the runway or
identify its centreline.
Taking into account the previous works and the RVR variable, we proposed
two works in the sense that fog prediction is treated as class intervals,
resulting in a binary and an ordinal problem, respectively. Both works are
located at the airport of Valladolid since its location is a key point in the for-
mation of fog during the cold months.
The first work is focused on the prediction of the RVR variable using data
collected from sensors with a 6-​hour time horizon prediction (Durán-​Rosal
et al., 2018). Fog events are considered when RVR is less than 1990 m. Thus,
the problem becomes a binary classification in which the output is 1 if there
is a fog event, and 0 otherwise. Given the unbalanced nature of the data set,
that is, the number of fog events was much lower than clear events, a multi-​
objective evolutionary algorithm (MOEA) was developed to train the struc-
ture and weights of ANNs, using PUs, sigmoid units (SUs) and radial base
functions (RBFs) as basis functions. The algorithm tried to optimize the overall
accuracy and minimum sensitivity to address the problem of unbalancing.
The best model obtained by MOEA was an ANN with a PU basis function.
This combination achieved a global accuracy of 82.24 per cent and the per-
centage of fog events correctly classified was 84.83 per cent. The resulting
ANN had only two neurons in the hidden layer, showing the high interpret-
ability of the model (see Figure 10.2), which was able to be explained in terms
of physical properties of fog formation.
In the second work (Guijo-​Rubio et al., 2018), three ordered classes were
considered, increasing the difficulty of the problem. The consideration of
these three classes was due to giving more precise information as a decision
support system: FOG (RVR in the range [0, 1000] m), MIST (RVR in the range
[1000, 1900] m) and CLEAR (RVR higher or equal to 1990 m). Besides, the
prediction was carried out using a 24-​hour schedule.
The work was based on the philosophy of autoregressive (AR) models, so
the model needed some previous values for the prediction of a given one. To
determine this number of previous values, three different sorts of window
were proposed: fixed window (FW), dynamic window (DW) based on label
change (DWLC) and DW based on variance change (DWVC). Since the
problem was addressed under the ordinal paradigm, the following algorithms
were proposed to solve it: proportional odds model (POM), support vector
machines (SVM), support vector ordinal regression considering explicit
constraints (SVOREX), support vector ordinal regression considering implicit
constraints (SVORIM) and kernel discriminant learning for ordinal regres-
sion (KDLOR).
ML Applications in Real-World Time Series Problems 169

FIGURE 10.2
Best ANN obtained by MOEA using PUs as basis functions. V and D are the velocity and wind
direction, respectively, T corresponds to the temperature, H represents the humidity, P is the
pressure, and RVR is the runway visual range previously defined.

The results showed that the combination of a FW and a DW with the algo-
rithm KDLOR is outstanding, producing the best results in terms of min-
imum sensitivity and AMAE (the average of ordinal classification error made
for each class). This methodology could lead to an improvement in the safety
and profitability of aviation operations at airports affected by low-​visibility
events.

10.2.4 Prediction of Other Convective Situations


As mentioned in Section 10.2.3., there is a great number of phenomena
interrupting the proper operation of airports. Not only do they cause endless
queues to the passengers but they also increase the workload of the operators
and workers (Cao and Cai, 2016). Thus, anticipating this sort of phenomena
has a significant impact nowadays. In this way, several proposals have been
presented in the literature, trying to improve our knowledge about con-
vective situations and aiming also to improve the current systems (Wang
et al., 2014). At the present time, most of the approaches are based on sta-
bility indices derived from the temperature and humidity, such as the lifted
index or the convective available potential energy index. Even though these
170 Machine Learning Algorithms and Applications in Engineering

indices have been proved to be useful in the prediction of conditions related


with convective situations, their performance can be increased by combining
them with ML techniques, which have also been proved to be excellent when
dealing with atmospheric issues (McGovern et al., 2017).
A convective cloud formation problem was considered by the author of the
chapter. Data collected from the airport of Madrid-​Barajas (Spain) for five
years (2011 to 2015) was used to present two novel approaches to the litera-
ture. The inputs considered in the problem were a combination of atmos-
pheric variables measured over the vertical with some popular stability
indices. The main goal of the study was the prediction of convective situ-
ations at 12 h time horizon. In this regard, two main proposals were made to
this problem: the first one (Guijo-​Rubio et al., 2020b) consisted in tackling the
problem from the nominal classification point of view. Each event was given
one of the following categories: clear (no convective clouds sighted), TCU
(cumulus congestus clouds sighted), CB (cumulonimbus clouds sighted)
and TS (thunderstop sighted). As can be imagined, this problem is highly
unbalanced given that there were more clear patterns than TCU, CB or TS.
Therefore, in this work, we analysed the performance of ANN trained by a
MOEA, aiming to achieve the best global performance without prejudicing
the smallest class (in this case, the TS). The results achieved demonstrated
that hybrid ANN (mixing basis functions in the hidden layer) are an excellent
approach for this kind of imbalanced problems.
On the other hand, the second proposal made in this field was tackled from
the ordinal point of view (Guijo-​Rubio et al., 2020c). Besides considering
that the classes followed an order relationship among them, in this work, an
oversampling method is previously applied as a preprocessing step, aiming
to balance the data set as much as possible. After that, ordinal methods
were applied to the resulting data set. SVORIM achieved the best results,
improving also the results achieved by terminal aerodrome forecasts (TAFs),
which are products available on the aerodrome that give concise statements
of the expected meteorological conditions, for almost all the categories.

10.3 Summary and Other Related Works


This chapter presents a set of TSDM applications using ML. The tasks the
chapter focuses on are reconstruction, segmentation and prediction, all of
them applied to time series data. The main applications presented, from the
reconstruction of missing values in wave height time series to the prediction
of low-​visibility situations in airports, provide the data in the form of time
series. The combination of time series data with the application of automatic
ML techniques to solve these real-​world problems has been proved to have
an excellent performance.
ML Applications in Real-World Time Series Problems 171

The authors of this chapter have addressed some other applications in real-​
world problems since time series can be found in a wide variety of fields, as
was mentioned above. A GA was proposed in combination with a likelihood-​
based segmentation procedure to automatically recognize financial patterns
in European stock market indexes (Durán-​Rosal et al., 2017a). The detection
and prediction of extreme waves using the retrieved time series obtained by
oceanographic buoys were also solved by applying a two-​stage algorithm
(Durán-​Rosal et al., 2017b). First, a segmentation algorithm, resulting from
an evolutionary approach hybridized with a likelihood-​based segmentation
involving a beta distribution, was used to detect the extreme events. Second,
an evolutionary ANN was proposed to predict them. Finally, in Guijo-​Rubio
et al. (2020d), evolutionary ANNs were successfully applied for the predic-
tion of global solar radiation at the radiometric station of Toledo (Spain)
using satellite-​based measurements.
Apart from this, there is a vast amount of works using time series in other
applications. Therefore, the potential of this type of temporal data is signifi-
cant due to the wide variety of solutions given to these problems by using
ML techniques. The future lies in the massive collection of data in real time,
its processing and the instantaneous generation of automatic ML techniques
to model and extract knowledge from them.

References
Abonyi, J., Feil, B., Nemeth, S., Arva. P. 2003. Fuzzy clustering based segmentation of
time-​series. In Advances in Intelligent Data Analysis V, 275–​285. https://​doi.org/​
10.1007/​978-​3-​540-​45231-​7_​26
Alzahrani, S. I., Aljamaan, I. A., Al-​Fakih, E. A. 2020. Forecasting the spread of the
COVID-​19 pandemic in Saudi Arabia using ARIMA prediction model under
current public health interventions. Journal of Infection and Public Health, 1(7),
914–​919. https://​doi.org/​10.1016/​j.jiph.2020.06.001
Belda, S., Pipia, L., Morcillo-​Pallarés, P., Rivera-​Caicedo, J. P., Amin, E., De Grave, C.,
Verrelst, J. 2020. DATimeS: A machine learning time series GUI toolbox for gap-​
filling and vegetation phenology trends detection. Enviromental Modelling and
Software 127, 104666. https://​doi.org/​10.1016/​j.envs​oft.2020.104​666
Balas, C., Koç, L., Balas, L. 2004. Predictions of missing wave data by recurrent
neuronets. Journal of Waterway, Port, Coastal, and Ocean Engineering, 130(5), 256–​
265. https://​doi.org/​10.1061/​(ASCE)0733-​950X(2004)130:5(256)
Bergot, T., Terradellas, E., Cuxart, J., Mira, A., Liechti, O., Mueller, M., Nielsen, N.W.
2007. Intercomparison of single-​column numerical models for the prediction of
radiation fog. Journal of Applied Meteorology and Climatology, 46, 504–​521. https://​
doi.org/​10.1175/​JAM2​475.1
Bhattacharya, B., Shrestha, D., Solomatine, D. 2003. Neural networks in reconstructing
missing wave data in sedimentation modelling. In Proceedings of the 30th IAHR
Congress, 500, 770–​778.
172 Machine Learning Algorithms and Applications in Engineering

Blázquez-​García, A., Conde, A., Mori, U., Lozano, J. A. 2020. A review on outlier/​
anomaly detection in time series data. ArXiv:2002.04236 [Cs], https://​arxiv.org/​
abs/​2002.04236
Bremnes, J. B., Michaelides, S. C. 2007. Probabilistic visibility forecasting using
neural networks. In Fog and Boundary Layer Clouds: Fog Visibility and Forecasting.
Springer, pp. 1365–​1381. https://​doi.org/​10.1007/​s00​024-​007-​0223-​6
Cao, Z., Cai, H. 2016. Identification of forcing mechanisms of convective initi-
ation over mountains through high-​ resolution numerical simulations.
Advances in Atmospheric Sciences, 33(10), 1104. https://​doi.org/​10.1007/​s00​
376-​016-​6198-​4
Casas-​Prat, M., Wang, X. L., Sierra, J. P. 2014. A physical-​based statistical method for
modeling ocean wave heights. Ocean Modelling, 73, 59–​75. https://​doi.org/​
10.1016/​j.oce​mod.2013.10.008
Chandra, R., Chand, S. 2016. Evaluation of co-​ evolutionary neural net-
work architectures for time series prediction with mobile application in
finance. Applied Soft Computing, 49, 462–​473. https://​doi.org/​10.1016/​
j.asoc.2016.08.029
Chmielecki, R. M., Raftery, A. E. 2011. Probabilistic visibility forecasting using
Bayesian model averaging. Monthly Weather Review, 139, 1626–​1636. https://​doi.
org/​10.1175/​2010​MWR3​516.1
Colabone, R. D. O., Ferrari, A. L., Vecchia, F. A. D. S., Tech, A. R. B. 2015. Application
of artificial neural networks for fog forecast. Journal of Aerospace Technology and
Management, 7, 240–​246. https://​doi.org/​10.5028/​jatm.v7i2.446
Cornejo-​Bueno, L., Casanova-​Mateo, C., Sanz-​Justo, J., Cerro-​Prada, E., Salcedo-​Sanz,
S. 2017. Efficient prediction of low-​visibility events at airports using machine-​
learning regression. Boundary-​Layer Meteorology, 165, 349–​370. https://​doi.org/​
10.1007/​s10​546-​017-​0276-​8
Dakos, V., Carpenter, S. R., Brock, W. A., Ellison, A. M., Guttal, V., Ives, A. R., Kefi,
S., Livina, V., Seekell, D. A., Van Nes, E. H, et al. 2012. Methods for detecting
early warnings of critical transitions in time series illustrated using simulated
ecological data. PloS One, 7(7), e41010. https://​doi.org/​10.1371/​jour​nal.
pone.0041​010
Durán-​Rosal, A. M., Hervás-​Martínez, C., Tallón-​Ballesteros, A. J., Martínez-​Estudillo,
A. C., Salcedo-​Sanz, S. 2015. Massive missing data reconstruction in ocean buoys
with evolutionary product unit neural networks. Ocean Engineering, 117, 292–​
301. https://​doi.org/​10.1016/​j.ocean​eng.2016.03.053
Durán-​Rosal, A. M., de la Paz-​ Marín, M., Gutiérrez, P. A., Hervás-​ Martínez, C.
2017a. Identifying market behaviours using European Stock Index time series
by a hybrid segmentation algorithm. Neural Processing Letters, 46(3), 767–​790.
https://​doi.org/​10.1007/​s11​063-​017-​9592-​8
Durán-​Rosal, A. M., Fernández, J. C., Gutiérrez, P. A., Hervás-​Martínez, C. 2017b.
Detection and prediction of segments containing extreme significant wave
heights. Ocean Engineering, 142, 268–​279. https://​doi.org/​10.1016/​j.ocean​
eng.2017.07.009.
Durán-​Rosal, A. M., Fernández, J. C., Casanova-​Mateo, C., Sanz-​Justo, J., Salcedo-​
Sanz, S., Hervás-​Martínez, C. 2018. Efficient fog prediction with multi-​objective
evolutionary neural networks. Applied Soft Computing 70, 347–​358. https://​doi.
org/​10.1016/​j.asoc.2018.05.035.
ML Applications in Real-World Time Series Problems 173

Fabbian, D., de Dear, R., Lellyett, S., 2007. Application of artificial neural network
forecasts to predict fog at Canberra international airport. Weather Forecast, 22,
372–​381. https://​doi.org/​10.1175/​WAF​980.1.
Fu, T. C. (2011). A review on time series data mining. Engineering Applications of
Artificial Intelligence, 24(1), 164–​181.
Guijo-​Rubio, D., Gutiérrez, P. A., Casanova-​Mateo, C., Sanz-​Justo, J., Salcedo-​Sanz,
S., Hervás-​Martínez, C. 2018. Prediction of low-​visibility events due to fog
using ordinal classification. Atmospheric Research, 214, 64–​73. https://​doi.org/​
10.1016/​j.atmos​res.2018.07.017.
Guijo-​Rubio, D., Durán-​Rosal, A. M., Gutiérrez, P. A., Troncoso, A., Hervás-​Martínez,
C. 2020a. Time-​series clustering based on the characterization of segment typ-
ologies. IEEE Transactions on Cybernetics (early access). http://​doi.org/​10.1109/​
TCYB.2019.2962​584
Guijo-​Rubio, D., Gutiérrez, P. A., Casanova-​Mateo, C., et al. 2020b. Prediction of con-
vective clouds formation using evolutionary neural computation techniques.
Neural Computing and Applications, 33, 13917–​13929. https://​doi.org/​10.1007/​
s00​521-​020-​04795-​w
Guijo-​Rubio, D., Casanova-​Mateo, C., Sanz-​Justo, J., Gutiérrez, P. A., Cornejo-​Bueno,
S., Hervás, C., Salcedo-​Sanz, S. 2020c. Ordinal regression algorithms for the ana-
lysis of convective situations over Madrid-​Barajas airport. Atmospheric Research,
236, 104798. https://​doi.org/​10.1016/​j.atmos​res.2019.104​798
Guijo-​Rubio, D., Durán-​Rosal, A. M., Gutiérrez, P. A., et al. 2020d. Evolutionary artifi-
cial neural networks for accurate solar radiation prediction. Energy, 210, 118374.
https://​doi.org/​10.1016/​j.ene​rgy.2020.118​374
Gunaydin, K. 2008. The estimation of monthly mean significant wave heights by using
artificial neural network and regression methods. Ocean Engineering, 35(14–​15),
406–​1415. https://​doi.org/​10.1016/​j.ocean​eng.2008.07.008
Karevan, Z., & Suykens, J. A. (2020). Transductive LSTM for time-​series prediction: An
application to weather forecasting. Neural Networks, 125, 1–​9.
Hamilton, J. D. 1994. Time Series Analysis, volume 2. Princeton University Press,
Princeton, NJ.
Ismail Fawaz, H., Forestier, G., Weber, J. et al. 2019. Deep learning for time series clas-
sification: a review. Data Mining and Knowledge Discovery 33, 917–​963. https://​
doi.org/​10.1007/​s10​618-​019-​00619-​1
Keogh, E., Chu, S., Hart, D., Pazzani, M. 2004. Segmenting time series: A survey and
novel approach. Data mining in Time Series Databases, 1–​21. https://​doi.org/​
10.1142/​978​9812​5654​02_​0​001
Lenton, T. M. 2011. Early warning of climate tipping points. Nature Climate Change, 1,
201–​209. https://​doi.org/​10.1038/​nclim​ate1​143
Liao, T. W. 2005. Clustering of time series data—​a survey. Pattern Recognition, 38(11),
1857–​1874. https://​doi.org/​10.1016/​j.pat​cog.2005.01.025
Londhe, S. 2008. Soft computing approach for real-​time estimation of missing wave
heights. Ocean Engineering, 35(11–​12), 1080–​1089. https://​doi.org/​10.1016/​
j.ocean​eng.2008.05.003
López, I., Andreu, J., Ceballos, S., de Alegría, I. M., Kortabarria, I. 2013. Review of
wave energy technologies and the necessary power-​ equipment. Renewable
and Sustainable Energy Reviews, 27, 413–​434. https://​doi.org/​10.1016/​
j.rser.2013.07.009
174 Machine Learning Algorithms and Applications in Engineering

Mahjoobi, J., Etemad-​ Shahidi, A., Kazeminezhad, M. 2008. Hindcasting of wave


parameters using different soft computing methods. Applied Ocean Research,
30(1), 28–​36. https://​doi.org/​10.1016/​j.apor.2008.03.002
McGovern, A., Elmore, K. L., Gagne, D. J., Haupt, S. E., Karstens, C. D., Lagerquist, R.
2017. Using artificial intelligence to improve real-​time decision-​making for high-​
impact weather. Bulletin of the American Meteorological Society, 98(10), 2073–​2090.
https://​doi.org/​10.1175/​BAMS-​D-​16-​0123.1
Marzban, C., Leyton, S., Colman, B. 2007. Ceiling and visibility forecasts via neural
networks. Weather and Forecasting, 22, 466–​479. https://​doi.org/​10.1175/​
WAF​994.1
Nikolaou, A., Gutiérrez, P. A., Durán, A., Dicaire, I., Fernández-​Navarro, F., & Hervás-​
Martínez, C. (2015). Detection of early warning signals in paleoclimate data
using a genetic time series segmentation algorithm. Climate Dynamics, 44(7),
1919–​1933.
Pérez-​Ortiz, M., Durán-​Rosal, A. M., Gutiérrez, P. A., Sánchez-​Monedero, J., Nikolauou,
A., Fernández-​Navarro, F., Hervás-​Martínez, C. 2019. On the use of evolutionary
time series analysis for segmenting paleoclimate data. Neurocomputing, 326–​327,
3–​14. https://​doi.org/​10.1016/​j.neu​com.2016.11.101
Rakthanmanon, T., Keogh, E. J., Lonardi, S., Evans, S. 2012. MDL-​based time series
clustering. Knowledge and Information Systems, 33(2), 371–​399. https://​doi.org/​
10.1007/​s10​115-​012-​0508-​7
Román-​Gascón, C. 2015. Radiation fog, gravity waves and their interactions with
turbulence in the atmospheric boundary layer. Ph.D. thesis, Universidad
Complutense de Madrid.
Román-​Gascón, C., Steeneveld, G., Yagüe, C., Sastre, M., Arrillaga, J., Maqueda, G.
2016. Forecasting radiation fog at climatologically contrasting sites: evaluation
of statistical methods and WRF. Quarterly Journal of the Royal Meteorological
Society, 142(695), 1048–​1063. https://​doi.org/​10.1002/​qj.2708
Setiawan, N., Venkatachalam, P., Hani, A. 2008. Missing attribute value prediction
based on artificial neural network and rough set theory. International Conference
on Biomedical Engineering and Informatics (MEI2008), 1, 306–​310. https://​doi.
org/​10.1109/​BMEI.2008.322
Thomson, R. E., Emery, W. J. 2014. Data Analysis Methods in Physical Oceanography.
Elsevier, San Diego, CA.
Tseng, V. S., Chen, C. H., Huang, P. C., Hong, T. P. 2009. Cluster-​based genetic seg-
mentation of time series with DWT. Pattern Recognition Letters, 30(13), 1190–​1197.
https://​doi.org/​10.1016/​j.pat​rec.2009.05.013
Wang, H., Luo, Y., Jou, B. J. D. 2014. Initiation, maintenance, and properties of con-
vection in an extreme rainfall event during SCMREX: Observational analysis.
Journal of Geophysical Research: Atmospheres, 199(23), 13–​206. https://​doi.org/​
10.1002/​2014J​D022​339
Wassmann, P., Lenton, T. M. 2012. Arctic tipping points in an earth system perspec-
tive. AMBIO, 41, 1–​9. https://​doi.org/​10.1007/​s13​280-​011-​0230-​9
Weigend, A. S. 2018. Time Series Prediction: Forecasting the Future and Understanding the
Past. Routledge, London.
Zamani, A., Solomatine, D., Azimian, A., Heemink, A. 2008. Learning from data for
wind-​wave forecasting. Ocean Engineering, 35(10), 953–​962. https://​doi.org/​
10.1016/​j.ocean​eng.2008.03.007
ML Applications in Real-World Time Series Problems 175

Zhao, H., Dong, Z., Li, T., Wang, X., Pang, C. 2016. Segmenting times series with
connected lines under maximum error bound. Information Sciences, 345, 1–​8.
https://​doi.org/​10.1016/​j.ins.2015.09.017
Zhou, J., Huang, Z. 2018. Recover missing sensor data with iterative imputing net-
work. In Workshops at the Thirty-​ Second AAAI Conference on Artificial
Intelligence.
newgenprepdf

11
Prediction of Selective Laser Sintering
Part Quality Using Deep Learning

Lokesh Kumar Saxena1,* and Pramod Kumar Jain2


1
Mechanical Engineering Department, J.M.I., New Delhi, India
2
Department of Mechanical and Industrial Engineering, Indian Institute of
Technology, Roorkee, India 247667
*Corresponding author. email:[email protected], [email protected]

CONTENTS
11.1 Introduction.............................................................................................. 177
11.1.1 Aim of the Chapter.................................................................... 180
11.2 Selective Laser Sintering Additive Manufacturing............................. 180
11.3 Machine Learning.................................................................................... 181
11.3.1 Data.............................................................................................. 182
11.3.2 Models......................................................................................... 182
11.3.3 Training....................................................................................... 182
11.3.4 Learning Type............................................................................ 183
11.4 Deep Neural Network Learning............................................................ 183
11.5 An Illustration Case................................................................................. 184
11.5.1 Dataset for the Chapter............................................................. 185
11.5.2 Deep Neural Network Parameters.......................................... 185
11.5.3 K-fold-Cross-validation for Training the Deep
Neural Networks....................................................................... 187
11.5.4 Overfitting.................................................................................. 187
11.5.5 Results and Discussion............................................................. 188
11.6 Conclusions............................................................................................... 189

11.1 Introduction
Manufacturing is the pillar of the economy that transforms raw materials into
products. Manufacturing is faced with a growing need for individualisation,

DOI: 10.1201/9781003104858-11 177


178 Machine Learning Algorithms and Applications in Engineering

quality, flexibility and efficiency. Numerous countries had shown a keen


interest in strategic plans to speed up the faster transformation from existing
mechanised, semi-​automatic or automatic manufacturing to smart manu-
facturing, for example, the United States with Advanced Manufacturing
Partnership and Germany with Industry 4.0. Smart manufacturing has
not been formally defined yet. But its core concepts have been recognised
by the researchers, namely, to monitor, control, integrate and optimise the
manufacturing processes employing the elements such as cyber-​physical
systems (CPS), Internet of Things (IoT), big data, advanced data analytics
and imparting intelligence to machines.[1–​4]
Traditionally, manufacturing, especially subtractive manufacturing,
removes the material from the surface of raw stock to make a product. In
contrast, additive manufacturing (AM) is an important modern digital
manufacturing industrial paradigm. It is eliciting a keen interest in manu-
facturing.[5–​14] It produces the product by adding the material layer by layer
employing three-​dimensional (3D) models. These models are made using
computer-​aided design software (CAD). AM ensures various benefits such
as producing complex shaped products with optimised structures in top-
ology and difficult to produce by conventional casting/​ forging, produ-
cing the novel material properties, for example, dislocation networks[15]
and decreasing the material waste to save the cost for industry. But AM
products also possess unique defects such as anisotropic microstructure in
the parallel and perpendicular and directions with respect to the printing
direction; porosity owing to gas entrapment and imperfect fusion; and dis-
tortion owing to residual stress resulting from faster temperature gradient
and very steep cooling rate.[16] Therefore, it is important to determine the
relationship between metallurgical properties of powder, AM process and
the mechanical properties and the microstructure of AM products. There
are various important parameters in the AM process responsible for the
characteristics of end products. For instance, the processing parameters for
selective laser sintering (SLS) are surrounding working temperature, laser
scanning speed, layer thickness, scanning mode, hatch distance, laser power
and interval time. The impacts of these parameters vary considerably with
respect to the quality of the produced products. But it is very difficult to
know the relationship between SLS parameters and product quality. The SLS
process needs no support structure to produce parts. It can process various
materials, for example, ceramics, nylon, polycarbonate, wax, nylon/​glass,
metal–​polymer powders and composite.[17] The dimensional accuracy of SLS
products is poor as compared to the conventional machining processes.[18]
Thus, it is important to enhance the accuracy of SLS process. Li et al. studied
the shrinkage parameters for SLS process.[19] John and Carl examined heat
transfer, the energy delivery and sintering process.[20] Yang et al. employed
the Taguchi method to examine the shrinkage compensation.[21] Masood et al.
examined the orientation of products.[22] Bai et al. examined the temperature
DL in Prediction of Selective Laser Sintering Part Quality 179

field for polymer–​molybdenum powder.[23] Arni and Gupta analysed perpen-


dicularity, parallelism and flatness tolerance.[24] Armillotta and Biggioggero
studied the impact of part orientation and layer thickness on surface finish.[25]
Shi et al. applied expert system with the neural network.[26] Product inaccuracy
results due to material contraction in the SLS.[27] Contraction produces stress
inside and deforms the parts. Power material properties and SLS parameters
have profound impact on the work contraction. Therefore, this chapter
aims at revealing the relationship among the SLS parameters with contrac-
tion. It is tiresome to relate parameters in SLS with part contraction by trad-
itional mathematical techniques in a short time. Besides the mathematical
models, data models have been employed in AM. These models are known
as machine learning (ML).[28,29] The primary benefit of this kind of model is
to learn automatically the relationship between the input characteristics and
output characteristics from previously available data without any physics-​
based equations. In various ML techniques, the neural network (NN) is the
most popular method. For instance, NNs are applied in areas such as nat-
ural language processing[30], voice recognition[31], computer vision[32] and
autonomous driving.[33] NN had created a profound effect on all supply
chain functions in industry encompassing product design, manufacturing
and distribution. The neural network had been applied for function approxi-
mation.[34] Three-​layered neural network is found to approximate nonlinear
continuous and non-​singular function.[35] NN exhibits a great ability to recog-
nise the underlying complicated patterns.
Deep learning has exhibited superior abilities to recognise the patterns in
the given data applying supervised learning and unsupervised extraction
of features.[36] Deep neural network learning is an extended version of the
multiple layered perception neural network including a greater number of
the hidden layers. It includes better efficient processing of information, more
effective computing power and nonlinear activation function. Deep learning
has many variants according to various network architectures and training
techniques to process various types of data, for example, auto-​encoder deep
network used for dimensional reduction application.[37] Deep neural network
was used to extract probabilistic feature application.[38] Convolutional neural
network is characterised by operations like pooling and convolution used for
application like pattern recognition and image processing. Recurrent neural
networks are used in applications like reasoning of the sequential data.
Having data of high amount and large diversity to train the deep neural net-
work, the deep neural network has the capability to pull the features auto-
matically. These can generate a causal relationship about the deep network
inputs with the network outputs. Deep neural network learning has become
a popular technique applied for natural language processing (NLP),[39] audio
processing,[40] computer vision (CV),[41] biomedical technology,[42] autonomous
driving[43] and intelligent applications.[44] But the use of deep neural network
learning approach is at a very early phase for the additive manufacturing yet.
180 Machine Learning Algorithms and Applications in Engineering

11.1.1 Aim of the Chapter


This chapter provides a review of the present progress to apply the NN algo-
rithm to AM. The authors consider that a chapter on deep learning applica-
tion to additive manufacturing will be useful to academicians, researchers
and production engineers to provide an understanding about deep learning.
Therefore, this chapter aims at helping to learn to apply the deep neural net-
work learning to predict product quality printed by additive manufacturing
with an exhibition of the model development and data analytics. In this
chapter, deep neural network is shown as the process modelling technique
to determine the relationship of the SLS parameters with the part contraction
ratio (shrinkage ratio is ratio of the intended value minus the actual value to
the intended value). SLS model, developed using deep neural network, has
the ability to forecast the amount of contraction resulting in part for a specific
set of the SLS parameters. Therefore, prior information about the dimensional
accuracy is necessary to take up the actual manufacturing of the product.
The remaining chapter is structured as follows. Section 11.2 provides an
introduction to AM technologies. Section 11.3 explains the fundamentals of
deep learning techniques. Section 11.4 demonstrates the approach to employ
the deep neural network learning to solve the problems in AM. Section 11.5
gives an illustration case. Section 11.6 provides the conclusions.

11.2 Selective Laser Sintering Additive Manufacturing


Powder bed fusion (PBF) is an AM process. It employs a laser thermal energy
source in the selective laser sintering additive manufacturing process to melt
and fuse selective regions of a powder bed. A typical SLS additive manu-
facturing process is shown in Figure 11.1. Many preparatory operations are
done to make an AM part such that the digital CAD model preparation, STL
file generation, STL file import in AM machine and data processing, material
powder preparation, the protective atmosphere generation and AM machine
warm up. In metal SLS process, the parts are often built on a metallic platform
to prevent the thermal strain-​developed deformation and for right orienta-
tion to reduce need of the support structures. Now, a thin material powder
layer is kept on the platform. Next, a cross-​section of sliced CAD model is
scanned employing a high-​power laser beam in order to melt partially and
fuse the material powder layer. Here, a laser beam spot travels along with a
scanning path. The scanning pattern is produced using the CAD model of the
product meant for production. Then, the platform travels down a distance of
the layer thickness.
Now, a powder supply platform is moved up in powder supply port to
transfer the powder over the build platform. A new material powder layer is
DL in Prediction of Selective Laser Sintering Part Quality 181

FIGURE 11.1
A typical SLS additive manufacturing process.

formed over the build platform by the roller in the AM machine. A powder
removal platform is moved down the powder removal port to remove the
excessive metal powder from the machine. These two processes—​the powder
layering and powder melting—​take place alternately till all products are
produced. Each layer is joined with the adjacent layers by the laser energy-​
fused material powder. In end, the parts on the build platform are extracted
out of the AM machine. The machine is now cleaned for the next AM part
production. Next, the parts on the build platform often undergo thorough
heat treatment to remove the thermal stress. Finally, the parts are cut from
the platform for performing the post-​processing actions like removing the
support structure and surface polishing.
Selective laser sintering (SLS) has the merits to recycle the leftover unpro-
cessed metal powder, time efficiency, energy efficiency and geometrical
freedom for product design. SLS is a promising AM manufacturing process
for various applications such as aerospace and automotive manufacturing.

11.3 Machine Learning
Machine learning is a promising area that employs existing data to predict/​
respond to the future data. Machine learning is used for computational
statistics, pattern recognition and artificial intelligence. Machine learning is
182 Machine Learning Algorithms and Applications in Engineering

vital for fields, for example, spam filtering, facial recognition and other areas
with no feasibility/​possibility to frame algorithms to do a task. Machine
learning is a technique to use machines, that is, computers with software, to
find insights from existing data. It also refers to the ability of the machines to
learn from the environment. Machines have been employed for human help
since the start of civilisation. Machine learning is a process of using an algo-
rithm to transform input data into parameters to interpret the future data.
Now, this section describes some key terms used in machine learning.

11.3.1 Data
Every learning technique is based upon data. A data set is employed to
impart training to the machine system. Data sets are gathered by people for
training. The data set may be of very large size. Machine control systems can
gather data using sensors from the system operation. It employs the data to
recognise the parameters or to train the machine system.

11.3.2 Models
Models are usually employed in the machine learning systems. The models
provide the mathematical framework for the machine learning systems.
A model is made by a person. It depends upon the human observations and
their experiences. For instance, a model of bus of length 1 and width 1 looking
from the top is of a rectangular shape with the size of a standard parking slot.
Models are often considered as man-​made to give a framework to machine
learning. But sometimes machine learning developed its models without any
man-​made structure.

11.3.3 Training
The machine learning system relates an input data to an output. It is required
training to perform this work. Similar to people, as the man needs training
to do the tasks, the machine also needs learning systems for their training.
Training is provided by feeding the machine system with a known input
and the corresponding known output to change the models or data in the
learning for relationship learning. Sometimes, it is similar to the curve fitting
technique or the regression technique. By having sufficient training data, the
machine system gains the ability to generate correct outputs corresponding
to new inputs. For instance, on feeding thousands of rat images to the face
recognition machine system, the images are generated to be of rats. Now, on
giving new rat images, the machine system will be able to identify these as
rats. If enough training data sets or the training data in terms of quantity or
variety is not provided, it may face a problem in identifying the rats.
DL in Prediction of Selective Laser Sintering Part Quality 183

11.3.4 Learning Type
This section describes types of learning as below.

Supervised Learning: Supervised learning is a learning in which a particular


data set for training is applied to train the machine learning system. Here, the
learning is known as supervised, since the training data sets are man-​made.
It does not show that people validate the results. The process to classify the
outputs of the machine system corresponding to given inputs is known as
labelling. In other words, one directly indicates whether the results are correct
or are there the desired outputs corresponding to every input set. Generating
the training data sets is very time-​consuming. It requires utmost care to ensure
the training data sets to give sufficient training to generate the correct results.
These must include the full variety of the inputs and desired outputs. After
the training, a set of test data is applied for the validation of results. When the
results of the test data aren’t of the desired quality, the test data is used as the
training set and the process is again repeated. For example, a person is trained
exclusively in mechanical engineering. If he was told to work on a civil engin-
eering problem, the results wouldn’t be as good as desired since the person
did not have the proper training for civil engineering.

Semisupervised Learning: In the semisupervised approach, some portion of


data set has the form of labelled training data sets and other data without
label. In reality, a small portion of the input data set is labelled data, since
labelling requires a skilled person. This small labelled data set is employed to
interpret the unlabelled data set.

Unsupervised Learning: Unsupervised learning is a learning type that does


not employ the training data sets. It is normally employed to identify the
patterns in data set without having any existing right answer. For instance,
on employing the unsupervised learning for training a face identification
machine system, the machine system may group the data in groups, some
groups out of those may be faces. Grouping techniques are often instances of
the unsupervised learning. The unsupervised learning enables the machine to
learn from the data the things without knowing in advance. It is an important
technique to find the hidden structures in data set.

11.4 Deep Neural Network Learning


Deep neural network learning technique evolved from traditional narrow
neural network approach. The major difference is that more than two hidden
layers are found in the deep neural network over the narrow neural network.
184 Machine Learning Algorithms and Applications in Engineering

The hidden layers are used to cast the input data on a space of multiple
dimensions. Here, the given input data may be examined applying various
perspectives . Greater the network hidden layer numbers, greater the probable
hidden patterns may be identified in a given data. But increasing layers creates
obstacles to train and perform the deep neural network such as the following.

(i) Severe vanishing gradient problem. Due increased depth of the neural
networks, it becomes difficult for some starting layers of the deep
neural network to get the forecasting error. It degrades the effectiveness
of the training considerably.
(ii) The risk of the overfitting considerably goes up from increased com-
plexity in the architecture of deep neural network, since it needs more
parameters for network training.
(iii) There is increased computation power required to deal with increased
complexity of the deep neural network and amount of training data .
But these problems had been solved up to some degree with advances
in deep neural network learning.

It is made possible by developments such as (i) a new structure of the neural


network to have the flow of the loss of information through the entire deep
neural network, and to refrain from the gradient disappearance phenomenon,
for example, deep neural network ResNet[45]; (ii) robust deep neural network
training employing the data sets of large volume alongside a new type acti-
vation functions to decrease the problem relating to overfitting; and (iii) use
of the faster computation resources like GPU[46]. Therefore, deep learning has
proved to be a modern technique for audio processing[26], natural language
processing[27] and computer vision[25].

11.5 An Illustration Case


In this section, an example is presented to predict the quality of the printed
part to illustrate the application of deep learning for additive manufacturing.
As described earlier, it is important to have good quality of the printed part,
and therefore, the authors propose to predict it. SLS is a complex additive
manufacturing process with many important parameters. Contraction ratio,
a measure of part quality, depends upon a number of SLS parameters. The
various important SLS parameters [47] are surrounding working temperature
(Te), scanning mode (Ms), layer thickness (lt), laser power (w), hatch spacing
(dt), scanning speed (v) and interval time (Ts). SLS parameters are to be mod-
elled by a vector X as the following.
DL in Prediction of Selective Laser Sintering Part Quality 185


X = [ w , lt , dt , v , Ts , Te , Ms ] (11.1)

The contraction ratio Y is to be modelled as function G with argument X:



Y = G (X ) (11.2)

The mathematical model of Equation (11.2) exhibits the effect of SLS


parameters on product quality, that is, the shrinkage ratio. When the function
G is represented by deep neural network Gn, Y is estimated as

Y = Gn ( X ) (11.3)

Here, the deep neural network is capable of expressing relationship of the con-
traction ratio with the SLS parameter. This chapter has considered a study[47]
for the demonstration of deep learning. Specimen material in this chapter
is taken as HBI, a composite of polystyrene in the study[47]. The shrinkage
ratio[47] is defined as

Y % = ([SD − SM ] SD ) × 100 (11.4)

where SD and SM are CAD model value and the measured value.

11.5.1 Dataset for the Chapter


The choice of design of the experiment is of paramount for deep learning. The
data set for experiments must have a wide representation of the problem [47].
In this chapter, there are the seven process parameters to examine as shown
in Table 11.1. The levels of each process parameter were chosen using some
references with experience[47].
The SLS process parameters and their levels are displayed for the
experiments in Table 11.1. GR-​200 analytical balance and AD-​1653 density
determination kit from A&D Company Ltd. (Tokyo, Japan) [47] is a suitable
equipment set to take measurements regarding the density of printed products
using SLS. SLS printed part’s density was determined corresponding to the
product weight in air medium.

11.5.2 Deep Neural Network Parameters


In this chapter, the implementation and training of the deep neural network
and its validation were conducted employing TensorFlow library [48]
in Python programming language [49]. The deep neural network had seven
input nodes for the seven SLS process parameters and the output node has
186 Machine Learning Algorithms and Applications in Engineering

TABLE 11.1
The SLS Parameters with Levels [47]

Parameters Levels Range

Layer thickness (mm) .1,.16,.2,.24 .1–​.24


Laser power (W) 8,8.5,9,10,11,11.5,12,12.5,14,15,18,20 8–​20
Scan speed (m/​s) 1300,1800,2000,2400,3000 1300–​3000
Hatch space (mm) .08,.1,.12,14,.15 .08–​.15
Interval time (s) 0,1,2,3 4 0–​4
Surroundings temperature of working (°C) 78,80,84,87,93,95 78–​95
Scan mode 1,2 1–​2

one node for one response characteristic, that is, SLS manufactured product’s
relative density. There was 10 hidden layers in the deep neural network. It is
found that due to the vast variations in values of seven input SLS parameters
and output as product quality characteristic, that is, contraction ratio, the
deep neural network learning process resulted in a deep neural network. So,
input characteristics as well as the output characteristics should be made in
normal form as below.
With mean of characteristic Xi and sample size N, Xmean is calculated as,



N
i =1
Xi
Xmean = (11.5)
N

Standard deviation of characteristic Xi is determined as,

− Xmean )
∑ (X
N 2
i =1 i
Xsd = (11.6)
N

Characteristic Xi is computed in normal form:

Xi − Xmean
X NormDeviate = (11.7)
Xsd

The rectified linear unit (ReLU), f(ZNormDeviate) =​max (0, ZNormDeviate) is an active
function, employed for dealing to move from one layer to the next layer
[50]
. This activation function is a very popular nonlinear function due to its
faster learning ability in deep neural networks having multiple layers of
neurons[51]. The sigmoid function, an active function, that is, f(ZNormDeviate) =​ 1/​
(1+​exp(-​ZNormDeviate)) was employed for deep neural network’s output layer
due to output nature of real value. For training of the deep neural network
DL in Prediction of Selective Laser Sintering Part Quality 187

in the operation of hyper parameter optimisation, the average square error


is employed as a loss function having Adam optimiser approach [52]. The loss
function is computed as

1  ∑ (YPredicted − Yreal ) 
N 2

Eod ( w ) =  i = 1 (11.8)
2 N 
 

where Yreal, Ypredicted and N are the actual magnitude of experimental output,
predicted magnitude of the model and grand number of SLS printed parts,
respectively.

K-​fold-​Cross-​validation for Training the Deep Neural Networks


11.5.3 
In order to train the deep neural network model, the data set is needed to
be divided into two parts: training data set and test data set. The data set
for training had a known output data. The deep neural network model
gains learning from this data to predict from other data. The data set for
test is employed to check the prediction accuracy of the deep neural net-
work model. But it is very difficult to divide the data set to represent the
required problem characteristics. Further, on division of the existing set of
data into validation and training data sets, the data set decreases consider-
ably the number of sample of data for use in learning of the model. This
is dependent upon a specific random selection of both data sets. So, k-​fold
cross-​validation is employed to solve problem specified here. For k-​fold
cross-​validation, the given original data set is divided at random in k-​equal-​
size data set subsamples. Among k subsets of the data sample, one single
subsample of data is marked as data set for the validation to validate deep
neural network model. Other (k –​1) data subsets are employed for training.
Now, cross-​validation procedure is performed k number of times (i.e., folds)
considering every k subsamples of data set employed one time only as the
testing data set for validation. Then, k number of results obtained from the
folds are taken to compute the average to create a single estimation[53]. The
merit of k-​fold technique is to employ all data sets to train and validate the
deep neural network. Every data set is employed to validate one time only. In
this chapter, the10-​fold cross-​validation approach was employed for training
the deep neural network.

11.5.4 Overfitting
Overfitting is a problem that usually happens in machine learning for very
accurate model with the training data set; however, it not very accurate with
validation data set. On the occurrence of overfitting, the neural network
model generally indulges in learning of the noise found within the training
188 Machine Learning Algorithms and Applications in Engineering

data set, rather than actual relationship among the parameters in the data
set. In order to refrain from overfitting, the dropout technique and weight
decay regularisation technique are employed. Weight decay technique is an
important technique to avoid weight to grow very high in value without real
requirements. This technique does so with the addition of a term in the loss
function to penalise the high weights [54,55] as in the equation below:


1
Ed ( w ) = Eod ( w ) + .λ.wi2 (11.9)
2

where λ is known as a regularisation parameter to control the extent of pen-


alisation of weights. W is known as the weight vector representing the entire
free variables/​parameters related to deep neural network. In the current
chapter, the regularisation parameter λ=​0.05 was employed. Drop out is
another important technique to avoid the overfitting phenomenon. In the drop
out mechanism, units are randomly removed from deep neural network’s
training phase[56–​58]. In the present chapter, the probability of keeping a unit in
deep neural network is taken as 0.7.

11.5.5 Results and Discussion


The deep neural network training aims at optimisation of the hyper
parameters of deep network model with minimisation of the loss function.
For this purpose, the weights are allocated randomly at the beginning. From
the given inputs and outputs data for the training, the deep neural network
estimates and makes comparison of its predicted outputs with the actual
outputs. The computer errors are sent back through the deep neural learning
system for adjusting the weights. The process of adjusting the weights is done
till the condition of minimal loss is achieved. On completion of the training,
the validation test dataset, a part of data set for the experiment, is employed
to validate the deep neural network. At last, an average error is computed to
obtain grand efficiency of deep neural network. For forecasting performance
level evaluation of the deep neural network, an average absolute error (AAE)
from Equation (12.10) is used.


∑ (Y )
− Yreal
N
i predicted
AAE = (11.10)
N

Figure 11.2 exhibits the 8-​fold-​cross-​validation strategy for deep neural net-
work having 440 epochs for every fold with 0.001 amount of learning rate.
Convergence for validation as well as training graphs shows that there is
no overfitting by the deep neural network. Further, the value of AAE is 1.53
for the training phase. The value of AAE is 1.54 for the validation phase.
DL in Prediction of Selective Laser Sintering Part Quality 189

FIGURE 11.2
The 8-​fold-​cross-​validation strategy for deep neural network.

This shows that there is similarity in the average AAE for training as well as
validation phase. The value of standard deviation for training phase is 0.21.
This lower standard deviation shows that deep neural network does not have
any tendency to change considerably with various training subset data. This
indicates a good forecasting ability of the deep neural network.

11.6 Conclusions
In this chapter, the quality of a product in terms of minimum shrinkage
ratio produced by selective laser sintering was predicted from important
SLS parameters such as surrounding working temperature, laser scanning
speed, layer thickness, scanning mode, hatch distance, laser power and
interval time. The relationship among SLS parameters with contraction
ratio is modelled using the deep neural network because the SLS variables
are considered to be multitudinous as well as nonlinear. In this chapter, the
machine learning system employed supervised deep learning including
process parameters as input characteristics and the quality of a product in
terms of minimum shrinkage ratio as output characteristic. Weight decay and
the dropout technique were employed to overcome the overfitting problem
found in deep neural network technique. The predicted outputs character-
istic was compared to the actual characteristic. The shrinkage ratio found
by the deep neural network can be employed to determine information for
shrinkage compensation in the SLS process.
190 Machine Learning Algorithms and Applications in Engineering

References
[1] Kang, H.S., J.Y. Lee, S. Choi, H. Kim, J.H. Park, and J.Y. Son et al. 2016. Smart
manufacturing: past research, present findings, and future directions. Int J Precis
Eng Manuf Green Technol, 3(January (1)):111–​28.
[2] Kusiak, A. 2017. Smart manufacturing must embrace big data. Nature, 544 (April
(7648)): 23–​25.
[3] Lee, J., B. Bagheri, and H.A. Kao. 2015. Cyber-​physical systems architecture for
industry 4.0 based manufacturing systems. Manuf Lett, 3 (January): 18–​23.
[4] Lee, J., H. Davari, J. Singh, and V. Pandhare. 2018. Industrial artificial intel-
ligence for industry 4.0-​ based manufacturing systems. Manuf Lett, 18
(October): 20–​23.
[5] Kumar, L. and P.K. Jain. 2010. Selection of additive manufacturing technology.
Adv Production Eng Mgmt J, 5(2): 75–​84.
[6] Kumar, L. and P. K. Jain (2006). Rapid prototyping: a review, issues and problems,
International Conference on CARs & FOF,VIT, India, 1: 126–​138.
[7] Kumar, L. and R. A. Khan (2004). Rapid design and manufacturing. Global
Conference on flexible System Management (GLOGIFT) March, JMI, Delhi,
1: 13–​15.
[8] Kumar, L., M. Shoeb, and A. Haleem (2020). An overview of additive manufac-
turing technologies. Studies in Indian Place Names, 40(10), 441–​450.
[9] Kumar, L., M. Shoeb, A. Haleem, and M. Javaid (2022). Composites in context
to Additive Manufacturing, CIMS-​2020-​International Conference on Industrial
and Manufacturing Systems, NIT Jalandhar, Punjab, India, Lecture Notes
on Multidisciplinary Industrial Engineering, Springer, Cham, 491–​503, ISBN
978-​3-​030-​73494-​7.
[10] Kumar L. and P.K. Jain (2022). Carbon conscious and artificial immune system
optimization modeling of metal powder additive manufacturing scheduling.
In Computational Intelligence for Manufacturing Process Advancements, eds. P.
Chatterjee, D.P. Željko-​Stević, S. Chakraborty, and S. Bhattacharyya, Taylor &
Francis, CRC Press (Accepted for publication by Editor).
[11] Lu, B, D. Li, and X. Tian. 2015. Development trends in additive manufacturing
and 3D printing. Engineering, 1(1):85–​89.
[12] Derby, B. 2015. Additive manufacture of ceramics components by ink jet printing.
Engineering, 1(1):113–​123.
[13] Gu, D., C. Ma, M. Xia, D. Dai, and Q. Shi. 2017. A multi scale understanding of
the thermodynamic and kinetic mechanisms of laser additive manufacturing.
Engineering, 3(5): 675–​684.
[14] Herzog, D., V. Seyda, E. Wycisk, and C. Emmelmann. 2016. Additive manufac-
turing of metals. Acta Mater, 117(15): 371–​392.
[15] Liu, L., Q. Ding, Y. Zhong, J. Zou, J. Wu, and Y.L. Chiu et al. 2018. Dislocation
network in additive manufactured steel breaks strength–​ ductility trade-​
off.
Mater Today, 21(4):354–​361.
[16] Gorsse, S., C. Hutchinson, M. Gouné, and R. Banerjee. 2017. Additive manu-
facturing of metals: a brief review of the characteristic microstructures and
properties of steels, Ti–​6Al–​4V and high-​entropy alloys. Sci Technol Adv Mater,
18(1): 584–​610.
DL in Prediction of Selective Laser Sintering Part Quality 191

[17] Santosa, E.C., M. Shiomia, K. Osakadaa, and T. Laoui. 2006. Rapid manufac-
turing of metal components by laser forming. Int J Mach Tools Manuf, 46(12–​
13): 1459–​1468.
[18] Chen, X., C. Wang, X. Ye, Y. Xiao, and S. Huang. 2001. Direct slicing from power
SHAPE models for rapid prototyping. Int J Adv Manuf Technol, 17(7): 543–​547
doi:10.1007/​s001700170156
[19] Li, X.S., M. Han, and Y.S. Shi. 2001. Model of shrinking and curl distortion for
SLS prototypes. Chin J Mech Eng, 12(8): 887–​889.
[20] John, D.W. and R.D. Carl. 1998. Advances in modeling the effects of selected
parameters on the SLS process. Rapid Prototyping J, 4 (2): 90–​96. doi:10.1108/​
13552549810210257
[21] Yang, H.J., P.J. Huang, and S.H. Lee. 2002. A study on shrinkage compensation of
SLS process by using the Taguchi method. Int J Mach Tools Manuf, 42(10): 1203–​
1212. doi:10.1016/​S0890-​6955 (02)00070-​6
[22] Masood, S.H., W. Ratanaway, and P. Iovenitti. 2003. A genetic algorithm for best
part orientation system for complex parts in rapid prototyping. J Mater Process
Technol, 139(3):110–​116. doi:10.1016/​S0924-​0136(03)00190-​0
[23] Bai, P.K., J. Cheng, B. Liu, and W.F. Wang. 2006. Numerical simulation of
temperature field during selective laser sintering of polymer coated molyb-
denum powder. Trans Nonferrous Met Soc China, 16 (3):603–​607. doi:10.1016/​
S1003-​6326(06)60264-​1
[24] Arni, R.K. and S.K. Gupta. 1999. Manufacturability analysis for solid freeform
fabrication. In: Proceedings of DETC 1999 ASME Design Engineering Technical
conference, Vegas, NV, 1–​12.
[25] Armillotta, A. and G.F. Biggioggero. 2001. Control of prototyping surface finish
through graphical simulation. In: Proceedings of the 7th ADM International
conferences, Grand Hotel, Rimini, Italy, 17–​24.
[26] Shi, Y., J. Liu, and S. Huang. 2002. The research of the SLS process optimization
based on the hybrid of neural network and expert system. In: Proceedings of the
International Conference on Manufacturing Automation, 409–​418.
[27] Zheng, H.Z., J. Zhang, S.Q. Lu, G.H. Wang, and Z.F. Xu. 2006. Effect of core–​
shell composite particles on the sintering behavior and properties of nano-​
Al2O3/​polystyrene composite prepared by SLS. Mater Lett, 60(9–​10): 1219–​1223.
doi:10.1016/​j.matlet. 2005.11.003
[28] Kohavi, R. and F. Provost. 1998. Glossary of terms. Mach Learn, 30(2–​
3): 271–​274.
[29] Géron, A. 2017. Hands-​on Machine Learning with ScikitLearn and Tensor Flow:
Concepts, Tools, and Techniques to Build Intelligent Systems. Boston, MA: O’Reilly
Media Inc.
[30] Devlin, J., M. Chang, K. Lee, and K.B. Toutanova. 2018. Pre-​training of deep bi-​
directional transformers for language understanding. arXiv, 1810:04805.
[31] Anusuya, M.A. and S.K. Katti. 2010. Speech recognition by machine-​a review.
arXiv, 1001–​2267.
[32] Krizhevsky, A., I. Sutskever, and G.E. Hinton. 2012. Image Net classification
with deep convolutional neural networks. In: F. Pereira, C.J.C. Burges, L. Bottou,
and K.Q. Weinberger, eds. Advances in Neural Information Processing Systems
25, Proceedings of Neural Information Processing Systems, 2012, December 3–​6,
Lake Tahoe, NV, 1097–​1105.
192 Machine Learning Algorithms and Applications in Engineering

[33] Ondruska, P. and I. Posner. 2016. Deep tracking: seeing beyond seeing using
recurrent neural networks, arXiv, 1602.00991.
[34] Mehrotra, P., J.E. Quaicoe, and R. Venkatesan. 1996. Speed estimation of induc-
tion motor using artificial neural networks. IEEE Trans Neural Netw, 6: 881–​886.
[35] Hornik, K., M. Stinchcombe, and H. White.1989. Multilayer feed forward
networks are universal approximators. Neural Netw, 2(5): 359–​366. doi:10.1016/​
0893-​6080(89)90020-​8
[36] Lecun, Y., Y. Bengio, and G. Hinton. 2015. Deep learning. Nature, 521 (May
(7553)): 436–​44.
[37] Hinton, G.E. and R.R. Salakhutdinov. 2006. Reducing the dimensionality of data
with neural networks. Science, 313 (July (5786)): 504–​507.
[38] Hinton, G.E., S. Osindero, and Y.W. Teh. 2006. A fast learning algorithm for deep
belief nets. Neural Comput, 18(7): 1527–​1554.
[39] Hirschberg, J. and C.D. Manning. 2015. Advances in natural language pro-
cessing. Science, 349 (July (6245)): 261–​266.
[40] Chan, W., N. Jaitly, Q. Le, and O. Vinyals. 2016. Listen, attend and spell: neural
network for large vocabulary conversational speech recognition. Proc. IEEE-​
ICASP, 4960–​4964.
[41] Liu, W., Z. Wang, X. Liu, N. Zeng, Y. Liu, and F.E. Alsaadi. 2017. A survey of
deep neural network architectures and their applications. Neuro Computing,
234(April): 11–​26.
[42] Mamoshina, P., A. Vieira, E. Putin, and A. Zhavoronkov. 2016. Applications of
deep learning in biomedicine. Mol Pharm, 13(May (5)):1445–​1454.
[43] Chen, C., A. Seff, A. Kornhauser, and J. Xiao. 2015. DeepDriving: learning
affordance for direct perception in autonomous driving. Proc. IEEE-​ ICCV,
2722–​2730.
[44] Covington, P., J. Adams, and E. Sargin. 2016. Deep neural networks for YouTube
recommendations. Proc. ACM—​RecSys, 191–​198.
[45] He, K., X. Zhang, S. Ren, and J. Sun. 2016. Deep residual learning for image rec-
ognition. Proc. IEEE-​CVPR, 770–​778.
[46] Sze, V., Y.H. Chen, T.J. Yang, and J.S. Emer. 2017. Efficient processing of deep neural
networks: a tutorial and survey. Proc. IEEE, 105 (December (12)): 2295–​2329.
[47] Rong-​Ji, W., L. Xin-​hua, W. Qing-​ding, and L. Lingling (2009). Optimizing pro-
cess parameters for selective laser sintering based on neural network and gen-
etic algorithm. Int J Adv Manuf Technol, 42: 1035–​1042.
[48] Abadi, M., P. Barham, J. Chen, Z. Chen, A. Davis, and J. Dean et al. 2016.
TensorFlow: a system for large-​scale machine learning. 12th USENIX Symp.
Oper. Syst. Des. Implement. (OSDI’ 16). https://​doi.org/​10.1038/​nn.3331
[49] Zelle, J. 2010. Python programming: an introduction to computer science.
https://​ doi.org/​10.2307/​2529413
[50] Glorot, X., A. Bordes, and Y. Bengio. 2011. Deep sparse rectifier neural
networks. AISTATS’ 11 Proc 14th Int Conf Artif Intell Stat, 15: 315–​ 23.
doi:10.1.1.208.6449
[51] LeCun, Y.A., Y. Bengio, and G.E. Hinton. 2015. Deep learning. Nature, 44: 436–​
444. https://​ doi.org/​10.1038/​nature14539
[52] Kingma, D.P., and J.L. Adam Ba. 2017. A method for stochastic optimization. Int
Conf Learn Represent. https://​doi.org/​10.1145/​1830​483.1830​503
DL in Prediction of Selective Laser Sintering Part Quality 193

[53] James, G., D. Witten, T. Hastie, and R. Tibshirani. 2013. An Introduction to


Statistical Learning, vol. 103. New York: Springer-​ Verlag. https://​doi.org/​
10.1007/​978-​1-​46147​138-​7
[54] Zur, R.M., Y. Jiang, L.L. Pesce, and K. Drukker. 2009. Noise injection for training
artificial neural networks: a comparison with weight decay and early stopping.
Med Phys, 36: 4810–​4818. https://​doi.org/​10.1118/​1.3213​517
[55] Krogh, A., and J.A. Hertz. 1992. A simple weight decay can improve general-
ization. Adv Neural Inf Process Syst, 950–​957. doi:https://​dl.acm.org/​citat​ion.
cfm?id=​2987​033
[56] Srivastava, N., G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov.
2014. Dropout: A simple way to prevent neural networks from overfitting. J
Mach Learn Res, 15:1929–​1958. https://​doi.org/​10.1214/​12-​AOS1​000
[57] Dahl, G.E., T.N. Sainath, and G.E. Hinton. 2013. Improving deep neural networks
for LVCSR using rectified linear units and dropout. ICASSP. IEEE Int. Conf.
Acoust. Speech Signal Process. Proc., 8609–​8613. https://​doi.org/​10.1109/​ICA​
SSP.2013.6639​346
[58] DeRosa, G.H., J.P. Papa, and X.S. Yang. 2018. Handling dropout probability
estimation in convolution neural networks using meta-​heuristics. Soft Comput,
22:6147–​6156. https://​doi.org/​10.1007/​s00​500-​017-​2678-​4
newgenprepdf

12
CBPP: An Efficient Algorithm for Privacy-​
Preserving Data Publishing of 1:M Micro
Data with Multiple Sensitive Attributes

Jayapradha Jayaram1,*, Prakash Manickam2, Apoorva Gupta3,


and Madhuri Rudrabhatla4
1
Department of Computer Science and Engineering, College of Engineering and
Technology, Faculty of Engineering and Technology, SRM Institute of Science and
Technology, SRM Nagar, Kattankulathur, Kanchipuram, Chennai, TN, India
2
Department of Computer Science and Engineering, College of Engineering
and Technology, Faculty of Engineering and Technology, SRM Institute of Science
and Technology, SRM Nagar, Kattankulathur, Kanchipuram, Chennai, TN, India
3
Department of Computer Science and Engineering, SRM Institute of Science
and Technology, Kattankulathur, Tamil Nadu, India
4
Department of Computer Science and Engineering, SRM Institute of Science and
Technology, Kattankulathur, Tamil Nadu, India
E-​mail: [email protected]; [email protected]; [email protected];
[email protected]

CONTENTS
12.1 Introduction.............................................................................................. 196
12.2 Related Works........................................................................................... 197
12.3 Contributions............................................................................................ 200
12.4 Correlation Attacks and Those Scenarios............................................. 200
12.5 Implementation of the Proposed CBPP Algorithm............................. 202
12.5.1 Generalization and Allocation of Batch Id............................. 204
12.6 CBPP Algorithm....................................................................................... 205
12.7 Evaluation................................................................................................. 207
12.8 Measurement of Information Loss........................................................ 207
12.8.1 Query Accuracy......................................................................... 207
12.9 Conclusion and Future Directions......................................................... 208

DOI: 10.1201/9781003104858-12 195


196 Machine Learning Algorithms and Applications in Engineering

12.1 Introduction
Preservation of the individuals’ privacy is an essential concern during and
after the data is transmitted to third person. In the era of digital world, data
owner organizations are discomforted by data privacy. Due to an increase
in the generation of data, the task of masking data without disclosing the
sensitive information of an individual has become a challenging task.
Organizations such as health sectors, pharmaceutical agencies and govern-
ment sectors often share their data with researchers and third parties for
various analysis. Therefore, the data publisher needs to take the responsi-
bility of preserving privacy during data publishing. The data publisher [1]
must be a trustworthy person. The data publisher needs to have a complete
knowledge about the privacy law and act before disclosing the data to the
data recipients. Data publisher should take the responsibility for the privacy
of the data and also the utility of the data. Data publisher ensures the data is
properly anonymized, so that the sensitive information of the original data
is unknown to the data recipient. However, the utility of the original data
should be preserved, so that the information needed is acquired correctly and
properly. The publisher needs to ensure that the proper model and techniques
have been applied on the original data before disclosing the data to the
recipient so that there is no clue for sensitive information. Two types of data
are of major concern before publishing the data: (i) individual privacy and
(ii) collective privacy. Revealing explicit identifiers such as name and id can
directly breach a particular individual privacy. To protect personal privacy,
the direct identifiers should be removed and the sensitive attributes related
to the particular individual need to be made private and anonymized. The
safeguarding of an individual privacy may not be adequate. Learning about
an individual’s sensitive information may also lead to inferring of the infor-
mation about a group of individuals. Therefore, sensitive knowledge about
the data set needs to be preserved. Preserving sensitive knowledge inferred
from the data set is termed as collective privacy preservation [2]. The data set
consists of two kinds of attributes termed as (i) sensitive attribute and (ii) non-​
sensitive attributes [3]. A distinctive care needs to be given for sensitive
attributes as they contain sensitive information about an individual. These
sensitive attributes should not be disclosed to the third party. Non-​sensitive
attributes in the dataset are projected for the purpose of analysis. The non-​
sensitive attributes are collectively termed as quasi-​identifier. Data publisher
discloses the non-​sensitive attributes to the third party. However, the non-​
sensitive attributes can collectively provide the personal information about
an individual if it is related with other external sources. Data anonymization
should be carried in a proper way by adopting various privacy methods and
models. The privacy-​preserving models and methods implemented on the
data set should balance between the privacy and utility. The basic notion of
CBPP for Privacy-Preserving Data Publishing 197

FIGURE 12.1
Privacy-​preserved data publishing.

privacy-​preserved data publishing is shown in Figure 12.1. The original data


set is transformed or anonymized by performing privacy-​preserving data
publishing models and methods. After anonymization, the data is protected
and ready to disclose it to the data recipient [4].

12.2 Related Works
Earlier, the researchers were concentrating on privacy-​ preserving data
publishing on a 1:1 dataset. An individual might have multiple records with
different sensitive attributes called a 1:M dataset in the real world. Consider
an individual who has cancer. A cancer disease comes with many side effects
such as weight loss, vomiting and hair loss. The same individual visits
different doctors in the hospital to be treated. So, each time he visits the hos-
pital, a record is registered in the hospital database. Such a scenario leads to
a 1:M dataset. When 1:1 privacy models are implemented on 1:M datasets, it
leads to various privacy breaches. Several techniques have been proposed for
a 1:1 dataset, such as slicing [5], Mondrian [6], suppression [7], clustering and
multi-​sensitive bucketization [8].
Another method called (k, km)-​anonymous [9] was proposed, which
divided the attributes into relational and transactional attributes. The limi-
tation of (k, km) –​anonymous is if the information loss for the relational
attribute is minimized, the information loss for the transactional attribute
would increase and vice-​ versa. An efficient approach (p,k)-​anglicization,
was proposed for the anonymization of multiple sensitive attributes. The
(p,k)-​anglicization provides the optimal balance between privacy and utility
[10]. A novel method called overlapped slicing was proposed for privacy-​
preserving data publishing. Overlapped slicing was implemented with mul-
tiple sensitive attributes. Overlapped slicing has proven to provide better
utility [11].
The anatomization prevents generalization of quasi-​identifiers, thus the
information loss is significantly less and results in higher privacy. However,
the anatomization method is efficient due to the publishing of multiple
198 Machine Learning Algorithms and Applications in Engineering

tables, and the complexity of the solution increases [12]. Anatomization


with slicing is an efficient method used for partitioning the data set with
less information loss. The method can preserve privacy, with numerous
sensitive attributes [13]. The (QI-​MHSA) generalization algorithm has been
proposed to anonymize the micro data with multiple sensitive attributes.
A vertical partitioning was applied and different models have been adopted
to anonymize the categorical and numerical attributes [14]. The Lsl-​diversity
model was proposed for multiple sensitive attributes. Along with Lsl-​diversity
model, three greed algorithms were also proposed to significantly reduce the
utility loss of the anonymized published data. Though the utility loss is stable
as the data grows, there is a slight increase in their runtime [15].
A model such as decomposition [16,17] was also proposed for privacy-​
preserving data publishing with multiple sensitive attributes. The sli-
cing and decomposition provide a solution for anonymization. However,
they lack utility and have significant information loss. When the slicing is
implemented without any anonymization method, it causes considerable
information loss; therefore, the suppression of the tuple was combined with
sliciVg [18]. (c,k)-​Anonymization, an improved method of (p,k)-​angliciza-
tion was proposed to thwart the fingerprint correlation attack. It improves
the one-​one correspondence relation in (p,k)-​anglicization to one-​to-​many
correspondence [19]. The traditional anonymization techniques such as k-​
anonymity, l-​diversity and t-​closeness resulted in optimal results for different
conditions [20]. Apart from data anonymization, hiding sensitive attributes
has also been implemented in health care, providing better security than
traditional and simple methods [21]. The traditional technique k-​anonymity
cannot be used on high-​dimensionality data as it results in a higher informa-
tion loss [22]. Various techniques have also been implemented and discussed
for data anonymization to reduce a high-​dimensional dataset into a smaller
dimensional data set [23,24]. Later, many researchers focused on 1:M dataset
and carried their work to provide an optimal balance between privacy and
utility. A new privacy model (k,l)-​diversity was proposed to address the re-​
identification risk in publishing of 1:M dataset. An efficient algorithm called
1:M generalization was proposed to preserve the data utility and privacy.
The (k,l)-​diversity model was proved to achieve the optimal balance between
privacy and utility. The INFORMS and YouTube data set has been used for
experimental analysis. However, the privacy model focused on single sensi-
tive attributes [25]. An algorithm named bucket-​individual multi-​sensitive
attribute bucket (QIAB-​IMSB) was proposed to achieve an optimal balance
between privacy and utility. The ultimate goal of QIAB-​IMSB algorithm is
anonymize the multi-​valued record with less utility loss and high privacy.
A vertical portioning was applied on the original data set and implemented k-​
anonymity on quasi-​identifier bucket and (k,l)-​diversity on sensitive attribute
buckets [26]. A model f-​slip was proposed to anonymize the 1:M dataset.
An effective anonymization method called “f-​slicing” was implemented to
newgenrtpdf
CBPP for Privacy-Preserving Data Publishing
TABLE 12.1
Sample Dataset Ts

Explicit
Identifier Quasi-​identifier Sensitive Attributes

Name Pid Sex Age Zipcode Disease Treatment Symptom Doctor Diagnostic Method

tp1
Avan 1 * 20-​30 142** HIV ART Infection John Elisa Test
tp2 Avan 1 * 20-​30 142** Influenza Medicine Fever Alice RITD Test
tp3 Avan 1 M 2* 142** Dyspepsia Antibiotics Abdominal Pain Victor Ultrasound
tp4 Becon 2 M 2* 142** Lung Cancer Radiation Weight Loss Alice MRI Scan
tp5 Becon 2 M,F 26 1420* Influenza Medicine Fever Alice RITD Test
tp6 Canty 3 M, F 28 1420* HIV Art Weight Loss John Elisa Test
tp7 Denny 4 M 25-​45 14249 Abdominal Cancer Chemotherapy Abdominal Pain Bob Chest Xray
tp8 Emy 5 F 25-​45 13084 Covid19 Antibiotics Fever Dave RT-​PCR Test
tp9 Emy 5 M,F 24-​45 13084 Asthma Medication Chest Tightness Alice Methacholine Challenge Test
tp10 Frank 6 M,F 24-​45 13064 Asthma Medication Shortness Of Breath Suzan Methacholine Challenge Test
tp11 Lisa 7 M, F 24-​45 13318 Lupus Medicine Joint Pain Jane Ana Test
tp12 Lisa 7 F 2* 1**** Myocarditis Medicine Abnormal Heartbeat Patrick ECG
tp13 Ram 8 M 2* 1**** Asthma Medication Shortness Of Breath Suzan Methacholine Challenge Test
tp14 Ram 8 M 2* 1**** Obesity Nutrition Control Eating Disorders Sana Body Mass Index

199
200 Machine Learning Algorithms and Applications in Engineering

anonymize the sensitive attributes. The f-​slip model thwarts five correlation
attacks such as (i) background knowledge attack, (ii) multiple sensitive cor-
relation attack, (iii) quasi-​identifier correlation attack, (iv) non-​membership
correlation attack, and (v) membership correlation attack [27]. As per the
study, the commonly used 1:M datasets are INFORMS and YouTube [25,26,27].
The chapter is organized as follows: Section 12.3 discusses the contribution
of the paper. Section 12.4 validates the various correlation attacks with the
scenarios. Section 12.5 discusses the implementation of the CBPP algorithm
and Section 12.6 presents the algorithm of CBPP. Section 12.7 describes the
evaluation and experimentation results that validate the effectiveness of the
CBPP algorithm. Section 12.8 concludes the work with future direction.

12.3 Contributions
The privacy-​preserving data publishing models should protect the microdata
with high privacy and less information loss. Though various algorithms have
been proposed to balance privacy and utility, the challenge the challenge
remains unsolved. The significant contributions of the work are as follows:

(1) A detailed study has been accomplished on various privacy-​preserving


data publishing models of the 1:M dataset.
(2) The anatomization is performed on the original microdata to partition
it into two tables.
(3) The quasi-​identifier and the sensitive attributes in the microdata are
carefully chosen and the distribution of the quasi-​identifier is computed.
(4) The generalization process of quasi-​identifier and allocation of batch
id is done using the CBPP algorithm.

12.4 Correlation Attacks and Those Scenarios


The proposed algorithm can resist three correlation attacks: (i) background
correlation attack, (ii) quasi-​
identifier correlation attack and (iii) non-​
membership correlation attack. Each correlation attack is discussed with a
scenario.

(1) Attacks using background knowledge correlation: The intruder can infer
the sensitive attributes of an individual if he possesses significant
background knowledge about the individual. Case scenario 1 explains
the background knowledge correlation attacks.
newgenrtpdf
CBPP for Privacy-Preserving Data Publishing
TABLE 12.2
Anonymity

Explicit
Identifier Quasi-​identifier Sensitive Attributes

Name Pid Sex Age Zip code Disease Treatment Symptom Doctor Diagnostic Method

tp1
Avan 1 M 27 14248 HIV ART Infection John Elisa Test
tp2 Avan 1 M 27 14248 Influenza Medicine Fever Alice RITD Test
tp3 Avan 1 M 27 14248 Dyspepsia Antibiotics Abdominal Pain Victor Ultrasound
tp4 Becon 2 M 26 14206 Lung Cancer Radiation Weight Loss Alice MRI Scan
tp5 Becon 2 M 26 14206 Influenza Medicine Fever Alice RITD Test
tp6 Canty 3 F 28 14207 HIV Art Weight Loss John Elisa Test
tp7 Denny 4 M 25 14249 Abdominal Cancer Chemotherapy Abdominal Pain Bob Chest Xray
tp8 Emy 5 F 44 13084 Covid19 Antibiotics Fever Dave RT-​PCR Test
tp9 Emy 5 F 44 13084 Asthma Medication Chest Tightness Alice Methacholine Challenge Test
tp10 Frank 6 M 45 13064 Asthma Medication Shortness Of Breath Suzan Methacholine Challenge Test
tp11 Lisa 7 F 24 13318 Lupus Medicine Joint Pain Jane Ana Test
tp12 Lisa 7 F 24 13318 Myocarditis Medicine Abnormal Heartbeat Patrick ECG
tp13 Ram 8 M 22 14421 Asthma Medication Shortness Of Breath Suzan Methacholine Challenge Test
tp14 Ram 8 M 22 14421 Obesity Nutrition Control Eating Disorders Sana Body Mass Index

201
202 Machine Learning Algorithms and Applications in Engineering

(2) Attack using quasi-​ identifiers correlation: The intruder can perform a
quasi-​identifier correlation attack by correlating the quasi-​identifier
values such as age, zip code, and gender to infer the individual sen-
sitive attributes value and the complete information of the individual.
Case scenarios 1 and 2 explain the quasi-​identifiers correlation attacks.
(3) Attacks using non-​membership correlation: The intruder can perform a
non-​membership correlation attack if he can infer the non-​existence of
an individual from the data set. The case scenarios 1, 2, and 3 explain
the non-​membership correlation attacks.
Scenario 1: If the intruder possesses the basic information of an indi-
vidual, he can gather sensitive information about that particular
individual. If the intruder knows that Emy is a female, age>40, from
zip code 13084, he can easily infer that Emy falls in either one of the
equivalence classes 4 and 5 from Table 12.2. If the intruder also has
strong background knowledge that Emy often suffers from breath-
lessness problems, the probability of inferring Emy’s record from
equivalence classes 4 and 5 is high.
Scenario 2: If the intruder possesses background knowledge about the
individual and the quasi-​identifier information, he can easily link
the sensitive attributes of an individual by using the quasi-​identifier
information. If the intruder knows that Becon is a male, age >25, zip
code 14206, then the intruder can easily infer that Becon falls in the
equivalence classes 2 and 3 from Table 12.2. With the strong back-
ground knowledge, he can quickly identify the values of the sensi-
tive attributes of Becon.
Scenario 3: If the intruder has strong background knowledge and
possesses the information of quasi-​identifier about the individual,
that is, Avan is a male, age <30, zip code 14248, also Avan is highly
infected with a deadly disease, then he can easily infer that the
records of Avan fall in equivalence classes 1 and 2 from Table 12.2.
So, the existence of the individual Avan is identified, which leads to
privacy breaches. The traditional methods and algorithms cannot
thwart the privacy breaches in the 1:M dataset.

12.5 Implementation of the Proposed CBPP Algorithm


Consider a 1:M dataset with multiple sensitive attributes. When the 1:M
dataset is generalized using traditional methods such as k-​anonymity, the
intruder can easily infer the sensitive values of the individual, as shown in
Table 12.2. The probability of occurrences of privacy breaches in Table 12.2 is
also discussed in the previous section.
CBPP for Privacy-Preserving Data Publishing 203

When traditional models and algorithms have been applied on the 1:M
dataset, it may cause privacy breaches. Therefore, an efficient CBPP algorithm
has been proposed to resist background knowledge attacks, non-​membership
correlation attacks and quasi-​identifier correlation attacks. CBPP algorithm
partitions the original data set into two tables: the quasi-​identifier table and
the sensitive attribute table. Both are linked together by a batch id, which is
further explained below.
Consider the original sample dataset Ts in Table 12.1. When the algorithm
CBPP is applied on the data set, it converts the original data set into two
tables. The original dataset Ts consists of multiple sensitive attributes for a
particular individual and shares the same quasi-​identifiers such as age, zip
code, and gender. Patient ID is just used for reference, and it will be removed
during the publishing of data. The records having same patient id are merged.
For example, the first three records of Table 12.1 with patient id 1 are merged.
(i.e., tp1Ս tp2Ս tp3). After merging the records of an individual, an aggregated
table Tsa is formed. Anatomization is performed on the aggregated dataset Tsa
and partitions the data set into two tables: (1) quasi-​identifier Tsq and (2) sen-
sitive attribute Tss. Let qd be the quasi-​identifier values of table Tsq and the set
of quasi-​identifier attributes are (qd1, qd2, qd3 … qdn). The quasi-​identifier table Tsq
comprises age, sex, and zip code.
Let sd be the values of the sensitive attributes of Tsa and the set of sensitive
attributes are (sd1, sd2, sd3 … sdn). The sensitive attributes are disease, treatment,
symptom, doctor, and diagnostic method. The quasi-​ identifier attribute
values are generalized and a batch id is allotted, as shown in Table 12.3. The
sensitive attribute table is formed as shown in Table 12.4 according to the
batch id. During the anonymization process, the data gets shuffled. Thus,
batch id, written as bid, is allocated to link the records in both quasi-​identifier
and sensitive attribute table.

TABLE 12.3
Quasi Identifier Table Tsq

Sex Age Zipcode Batch Id

M 25-​45 13000-​15000 2
M 25-​45 13000-​15000 2
F 25-​45 13000-​15000 2
M 25-​45 13000-​15000 2
F 41-​60 13000-​15000 3
M 41-​60 13000-​15000 3
F 15-​24 13000-​15000 1
M 15-​24 13000-​15000 1
204 Machine Learning Algorithms and Applications in Engineering

TABLE 12.4
Sensitive Attribute Batch Table T

Diagnosis
Disease Treatment Symptom Physician Method Batch id

Influenza, HIV, Art, medicine, Infection, fever, John, ELISA Test, 1


asthma, medication, shortness of Alice, RITD Test,
obesity, lupus, nutritional breath, eating Suzan, methacholine
myocarditis control disorder, Sana, challenge Test,
joint pain, Jane, body mass
abnormal Patrick index, Ana
heartbeat Test, ECG
HIV, influenza, Art, medicine, Infection, fever, John, ELISA Test, 2
dyspepsia, radiation, abdominal Alice, RITD Test,
lung cancer, chemotherapy, pain, weight Victor, Ultrasound,
abdominal medication loss, chest Bob, MRI scan,
cancer, tightness, Suzan, chest X-​ray,
asthma shortness of methacholine
breath challenge test
Phthisis, Antibiotics, Fever, chest David, Molecular 3
asthma, medication, tightness, Suzan, diagnostic
influenza, medicine, shortness Alice, methods,
HIV, radiation, art, of breath, John, methacholine
dyspepsia, chemotherapy, infection, Victor, challenge test,
abdominal medication abdominal Bob, RITD test,
cancer pain Dave ELISA test,
ultrasound,
chest X-​ray,
RT-​PCR test

12.5.1 Generalization and Allocation of Batch Id


The generalization process replaces the original value into a particular range
of values (e.g.) if an individual’s age is 25, then it can be replaced in the
interval of 20–​30. The allocation of batch id prevents the privacy breaches as
the tables to be published only with batch id. The batch id, bid, is allotted by
using the combinations of quasi-​identifiers.
Assume that we generalize the age as 25–​30, 30–​35, and so on. So, the batch
id is allotted to each group as 1, 2 up to n if there are n distributions in a
particular attribute. In the sample data set, three different quasi-​identifiers
are present. To find the total number of possible combinations, let’s say that
values of sex age and zip code are distributed in three groups. Therefore, the
total number of batch id combinations possible is 2*3*1+​3 =​9. Out of these
9 possible combinations, any combination can be used. When the number
of combinations is high, it can be decided by the user. Therefore, it becomes
challenging to find the individual in the possible combination. Hence, batch
id helps in preserving the privacy of the data. The complete architecture of
the CBPP algorithm is shown in Figure 12.2.
CBPP for Privacy-Preserving Data Publishing 205

FIGURE 12.2
Architectural diagram of the CBPP algorithm.

12.6 CBPP Algorithm
The algorithm has been divided into two parts and each part is elaborated.
In algorithm 1, the original data set is passed as an input argument in
line 1. In line 2, the multi-​records of an individual are merged using the
group by function such that each individual has only one entry. In line 3,
anatomization has been performed to partition the original table Ts into two
tables: (a) quasi-​identifiers (Tsq) and (b) sensitive attribute table (Tsa). In line
4, the quasi-​identifier is identified as sex, age, and zip code. The new list for
age and zip code are created in lines 5 and 6. In line 7, a new list has been
created for batch id. If the value for age ranges from 15 to 60, the optimal
number of distributions is 3, that is, 15–​24, 25–​44, and 45–​60. From lines 8 to
10, the quasi-​identifier age has been generalized. For example, if the person’s
age is 23, then the value 15–​24 will be appended to the list. Similarly, in lines
11–​13, the quasi-​identifier zip code, distributions are formed and values are
appended into the zip code list stated in lines 14 and 15. In the sample data
set of the work, the zip code ranges from 13000 to 15000. So, in Table 12.3, the
distribution of zip code ranges from 13000 to 15000 for all the individuals.
According to the algorithm, the total number of batch id combinations pos-
sible is 2*3*3+​3 =​21, whereas the original sample data set of the work has the
total number of batch id combinations possible as 2*3*1+​3 =​9.
In algorithm 2, sensitive attributes are arranged according to the batch id
and the quasi-​identifier that has been used. In line 1, the anatomized sensitive
attribute table has been passed as an argument. From lines 2 to 4, three lists of
each sensitive attribute have been created. Specifically, three lists are created
because the number of distributions that have been made in the quasi-​identifiers
is 3 and that quasi-​identifier in return gives existence to batch id. In line 5, iterate
the values according to age and each value is sent to the list corresponding to
206 Machine Learning Algorithms and Applications in Engineering

the batch id. A similar procedure is performed for all the sensitive attributes and
the sensitive values are clubbed together according to batch id.

Algorithm 1: Creating quasi-​identifier table from original aggregated table


1. Input(Ts)
2. Tsa ← Ts.groupby(patient_​id)
3. anatomization(Tsa)
4. Tsq ← Ts [“sex”, “age”, “zipcode”]
5.  age_​class ← []
6.  zip_​class ← []
7.  Batch_​id ← []
8.  For x in range (length(Tsq)
  a ← Tsq.loc(i, age)
  If(a > 15) && (a < 25)
  age_​class.append(15-​24)
 bid.append(1)
9.  If(a>=​25) && (a<=​40)
  age_​class.append(25-​40)
 bid.append(2)
10.  If(a>=​41) && (a<60)
age_​class.append(41-​60)
bid.append(3)
11.  for y in range(length(Tsq)):
   b ← df_​gen.loc[i, ‘ZIPCODE’]
  if (b < 10000) and (b > 13000):
   zipcode_​class.append(“10000-​13000”)
12.  elif (b >=​13000) and (b < 16000):
   zipcode_​ class.append(“13000-​16000”)
13. else:
   zipcode_​ class.append(“16000-​19000”)
14. Tsq.drop(age, zipcode)
15. Tsq (age_​class,zip_​class)

Algorithm 2: Creating sensitive attribute from aggregated table


1) Input(Tsa)
2) sd1 _​list ← []
3) sd2_​list ← []
4) sd3_​list ← []
5)  for i in range(length(df)):
  age ← df.loc[i, age]
   if (age >=​15) and (age < 26):
   # For Disease
   d ← df.loc[i, ‘disease’]
  slist ← s.split(‘,’)
    for dis in slist:
     if dis not in dis_​1:
     dis1.append(dis)
6) END
CBPP for Privacy-Preserving Data Publishing 207

12.7 Evaluation
The CBPP algorithm is implemented in Python and the experiments are
conducted in the machine that runs on Windows 10, 8 GB RAM, 1 TB storage
and 128 GB SSD. The CBPP algorithm is performed on the real-​world 1:M
dataset INFORMS. There are 2, 30,231 records in INFORMS dataset. After
grouping of records and removal of duplicates, the size of the dataset is
40,126. The birth year, month, sex, and race are chosen as quasi-​identifiers
and education year, income, and poverty line are taken as sensitive attributes.
The information loss is measured using query accuracy.

12.8 Measurement of Information Loss


12.8.1 Query Accuracy
Since the information loss needs to be measured from both the tables, (i) quasi-​
identifier and (ii) sensitive attribute, usual methods of utility measurement
cannot be performed. Hence, we use query accuracy for the calculation utility
of the algorithm. The query accuracy measures the information loss with the
result of aggregate queries.


Query Error =
∑ (QI ) − ∑ ( Org )
(12.1)
count ( Org )

The query error measures the COUNT queries executed on the micro-​table
data set and original data set for measuring information loss. In this case, the
total number of batch ID combinations possible is 9. Out of these 9, only one
particular combination is the key that links the quasi-​identifier table and the
sensitive batch table. Hence, batch id is the first line of protection. The prob-
ability that the attacker successfully finds the perfect combination becomes 1/​
9 =​0.11, which is very low. The combinations of batch id can be increased by
making more distributions in age and sex, further decreasing the probability
of identifying the batch id combination. In Figure 12.3, it is clearly shown that
information loss is inversely proportional to the number of values. Therefore,
information loss would decrease with the increase in the number of records.
However, in k-​anonymity, information loss increases as the dimensionality
increases [18], as shown in Figure 12.3. The CBPP algorithm provides better
utility when compared with the traditional method, k-​anonymity.
Consider that, in some cases, the attacker can get through the batch ID.
Then the intruder further has to make proper combinations of all the shuffled
208 Machine Learning Algorithms and Applications in Engineering

FIGURE 12.3
Information loss.

FIGURE 12.4
Probability of identification of quasi-​identifier combination versus number of distributions.

sensitive values to obtain the proper and complete details about the individ-
uals. The relation between the distributions and the probability to crack the
combination is given in Figure 12.4.

12.9 Conclusion and Future Directions


The study presents an efficient algorithm for privacy-​
preserving data
publishing of the 1:M data set. The proposed CBPP algorithm addresses
CBPP for Privacy-Preserving Data Publishing 209

the solution for three correlation attacks such as background correlation


attack, quasi-​identifiers correlation attack, and non-​membership correlation
attack. The case scenarios in the chapter explain that the traditional method
cannot be applied to the 1:M dataset. The probability of privacy breaches
was determined and the information loss is measured and compared with
anonymity. Experiments were carried out on the real-​
k-​ world dataset
INFORMS. The query accuracy is used as a measure for information loss to
show that CBPP is providing less information loss compared to the traditional
method. Still, there are contents to be worked on for future directions. The
probability of identifying the combinations upon the distribution of quasi-​
identifier using batch id should be implemented on a higher dimension data
set. Also, the study focuses on static QID, whereas the quasi-​identifier can
also change as time changes.

References
[1] Rashid A. and Yasin N., Privacy preserving data publishing: review. International
Journal of Physical Sciences, 10(7), 239 (2015).
[2] Sheppard, E., Data Privacy is a collective concern, 2020. https://​med​ium.com/​
open​data​char​ter/​data-​priv​acy-​is-​a-​col​lect​ive-​conc​ern-​8ebad​29b2​5ce
[3] Vasudevan L., Sukanya S., and Aarthi N., Privacy preserving data mining
using cryptographic role based access control approach, Proceedings of the
International MultiConference of Engineers and Computer Scientists, Hong
Kong, March 19–​21, 2008, p. 474.
[4] Churi Prathamesh, P. and Pawar Ambika, V., A systematic review on privacy pre-
serving data publishing techniques, Journal of Engineering Science and Technology
Review, 12(6), 17–​25 (2019).
[5] Li, T., Li, N., Zhang, J., and Molloy, I. Slicing: a new approach for privacy-​
preserving data publishing, IEEE Transactions on Knowledge and Data Engineering,
24(3), 561–​574 (2010).
[6] Wantong Zheng, Zhongye Wang, Tongtong Lv, Yong Ma, and Chunfu Jia,
K-​anonymity algorithm based on improved clustering, International Conference
on Algorithms and Architectures for Parallel Processing, 11335, 426–​476 (2018).
[7] Elanshekhar, N. and Shedge, R. An effective anonymization technique of big
data using suppression slicing method, International Conference on Energy,
Communication, Data Analytics and Soft Computing (ICECDS), 2500–​ 2504
(2017).
[8] Radha, D. and Valli Kumari Vatsavayi, Bucketize. Protecting privacy on multiple
numerical sensitive attribute, Advances in Computational Sciences and Technology,
10(5), 991–​1008 (2017).
[9] Puri, V., Sachdeva, S., and Parmeet Kaur, Privacy preserving publication of
relational and transaction data: Survey on the anonymization of patient data,
Computer Science Review, 32, 45–​61 (2019).
210 Machine Learning Algorithms and Applications in Engineering

[10] Anjum, A., Ahmad, N., Malik, U.R., Zubair, S., and Shahzad, B., An efficient
approach for publishing micro data for multiple sensitive attributes, The Journal
of Supercomputing, 74, 5127–​5155 (2018).
[11] Widodo, Budiardjo, E.K., and Wibowo, W.C. Privacy preserving data publishing
with multiple sensitive attributes based on overlapped slicing. Information, 10,
362 (2019).
[12] Lin Yao, Zhenyu Chen, and Xin Wang, Dong Liu, and Guowei Wu, Sensitive label
privacy preservation with anatomization for data publishing, IEEE Transactions
on Dependable and Secure Computing, 18(2), 904–​917 (2019).
[13] Susan, V.S. and Christopher, T. Anatomisation with slicing: a new privacy pres-
ervation approach for multiple sensitive attributes. SpringerPlus 5, 964 (2016).
[14] Jayapradha, J., Prakash, M., and Harshavardhan Reddy, Y., Privacy preserving
data publishing for heterogeneous multiple sensitive attribute with personalized
privacy and enhanced utility, Systematic Reviews of Pharmacy, 11(9), 1055–​1066
(2020).
[15] Yuelei Xiao and Haiqi Li, Privacy preserving data publishing for multiple sensi-
tive attributes based on security level, Information, MDPI, 11, 1–​27 (2020).
[16] Das, D. and Bhattacharyya, D.K., Decomposition: improving l-​diversity for mul-
tiple sensitive attributes, International Conference on Computer Science and
Information Technology 403–​412 (2012).
[17] Yang Ye, Liu Yu, Chi Wang, Depang Lv, and Jianhua Feng, Decomposition: privacy
preservation for multiple sensitive attributes. International Conference on
Database Systems for Advanced Applications, 486–​490 (2009).
[18] Kiruthika, S. and Mohamed Raseen, M., Enhanced slicing models for preserving
privacy in data publication, International Conference on Current Trends in
Engineering and Technology, IEEE, 1–​8 (2013).
[19] Khan, R., Tao, X., Anjum, A., Sajjad, H., Khan, A., and Amiri, F. Privacy pre-
serving for multiple sensitive attributes against fingerprint correlation attack
satisfying c-​diversity. Wireless Communications and Mobile Computing, 1–​18
(2020).
[20] Bennati, S. and Kovacevic, A., Privacy metric for trajectory data based on k-​
anonymity, l-​diversity and t-​closeness, (2020), https://​arxiv.org/​pdf/​2011.0921​
8v1.pdf
[21] Pika, A., Twynn, M., Budiono, S., Ter Hofstede, A.H.M, van der Aalst, wil MP,
and Reijers, H.A. Privacy-​preserving process mining in healthcare, International
Journal of Environmental Research Public Healthcare, 17(5), 1612 (2019).
[22] Aggarwal, C.C. On k anonymity and curse of dimensionality, 31st International
Conference on Very Large Databases, 433–​460 (2018).
[23] Bild, R., Kuhn, K.A., and Prasser, F., Better safe than sorry—​implementing reli-
able health data anonymization, Student Health Technology Inform, 270, 68–​72
(2020).
[24] Wang, R., Zhu, Y., Chen, T., and Chang, C. Privacy-​preserving algorithms for
multiple sensitive attributes satisfying t-​closeness. Journal of Computer Science
Technology 33, 1231–​1242 (2018).
[25] Qiyuan Gonga, Junzhou Luo, Ming Yang, Weiwei Ni, and Xiao-​ Bai Li,
Anonymizing 1:M microdata with high utility, Knowledge-​Based Systems, 1–​12
(2016).
CBPP for Privacy-Preserving Data Publishing 211

[26] Jayapradha, J. and Prakash, M. An efficient privacy-​preserving data publishing


in health care records with multiple sensitive attributes, 6th International
Conference on Inventive Computation Technologies, 623–​629 (2021).
[27] Jayapradha, J. and Prakash, M, f-​slip: An efficient privacy-​preserving data,
publishing framework for 1:M microdata with multiple sensitive attributes, Soft
Computing, pp. 1–​18, (2021).
newgenprepdf

13
Classification of Network Traffic on
ISP Link and Analysis of Network
Bandwidth during COVID-​19

V. Ajantha Devi,1 Yogendra Malgundkar,2 and Bandana Mahapatra2


1
Research Head, AP3 Solutions, Chennai, India
2
Symbiosis Skills and Professional University, India
E-​mail: [email protected]

CONTENTS
13.1 Introduction.............................................................................................. 214
13.2 Methodology............................................................................................. 214
13.2.1 Network Topology..................................................................... 217
13.2.2 Data Cleaning and Preprocessing........................................... 218
13.2.3 Data Visualization with Tableau.............................................. 218
13.3 Traffic Classification Using Classification Algorithms....................... 220
13.3.1 Comparison of Classifiers......................................................... 222
13.4 Bandwidth Requirement Prediction Using Time Series Model
(ARIMA).................................................................................................... 223
13.4.1 Feature Extraction Selection..................................................... 223
13.4.2 Auto-Regressive Integrated Moving Average Model
(ARIMA)..................................................................................... 224
13.4.3 Time Series Data: Trends and Seasonality.............................. 225
13.4.4 Time Series Data: Seasonal Patterns........................................ 227
13.5 Implementing the ARIMA Model.......................................................... 230
13.5.1 Building ARIMA Model............................................................ 231
13.5.2 Printing the Forecasted Values of Bandwidth....................... 232
13.5.3 Model Evaluation...................................................................... 234
13.6 Conclusion: Business Benefits of the Models....................................... 234
13.6.1 Network Traffic Classification................................................. 234
13.6.2 Time Series Prediction............................................................... 235

DOI: 10.1201/9781003104858-13 213


214 Machine Learning Algorithms and Applications in Engineering

13.1 Introduction
Nowadays, ISP networks are more complex than before, the number of
applications that run on clients/​servers are increasing leading to a network
resource management problem. The IT professionals tasked with maintaining
the network are facing serious challenges in determining which applications
consume the resources or degrade network performance. Analyzing the per-
formance metrics collected from clients/​servers using traditional methods
is not enough anymore to make a correct decision whether an application
consumes more network resources. The improvement of the network per-
formance is not an easy task, and even a network bandwidth upgrade might
not be an optimal solution to solve the problem of high network utilization.
In this study, taking advantage of machine learning, business intelligence
and data analytics to compare and show the benefits of using those techniques
in analyzing the network performance metrics, performance metrics from a
real ISP network are collected. A cleaned data with performance metrics will
be analyzed to find the pattern and correlation to give a better understanding
of the applications and network performance. The data set will be used to
apply machine learning algorithms to predict future network performance
under certain conditions. The implementation of those techniques in future
will cut the cost of running a network and reduce investigation time when-
ever a problem occurs and make IT professional’s life much easier.
Due to the COVID-​19 situation, almost 100% workforce across organizations
and various businesses were working from home/​remotely. This results in
heavy utilization of internet bandwidth at the ISP side. As a result, it’s crit-
ical for ISPs to understand and classify the many types of network traffic
that pass through their network. Internet service providers can use this clas-
sification to successfully control the network performance of their Internet
lines and deliver a high-​quality service to their consumers, leading to high
customer satisfaction.
Network traffic classification is also significant for network security and for
determining intrusion detection, QoS, and other features. Forecasting internet
bandwidth demand would also greatly assist ISPs in efficiently planning their
network resources. ISPs are interested in centralized measures and detecting
problems with specific customers before the they complain about the difficul-
ties, and if possible, before the consumers discover the problems at all.

13.2 Methodology
An experiment will be conducted out to ascertain the effects of applications
on resource consumption in the networks by extracting performance metrics.
The data is analyzed using machine learning and data analytics techniques,
Classification of Network Traffic on ISP Link during COVID-19 215

in addition to finding any correlation that exists between those metrics. The
performance metrics are extracted from the client as well as the server. The
clients are accessing files or web applications over the internet links.
The methodology is outlined as follows:

• Data collection
• Parsing the data—​data cleaning, scaling, and normalization as required
• Generating graphs—​to find trends and patterns in utilizations, find
correlations
• Data modelling using ML techniques to predict the class of network
traffic
• Predictive modelling to forecast the bandwidth requirement
• Analysis technique: would include Machine Learning classifiers.

Data Collection Method: This is the most crucial and first step, which
comprises data collection. The real-​time network traffic is recorded in this
step. There are several programs for capturing network traffic; however, the
TCP dump utility can capture real-​time network information. The solar wind
packet capturing and analysis program is utilized to gather network traffic.
Application traffic such as WWW, DNS, FTP, P2P, and Telnet is recorded.
Random network connection details were captured for the duration of one
year as shown in Figure 13.1.

FIGURE 13.1
Random network connection captured in mid-​July 2019 to mid-​July 2020.
216 Machine Learning Algorithms and Applications in Engineering

FIGURE 13.2
Network traffic data is collected for one of the links from the ISP.

Network traffic data is collected for the defined duration for one of the
links from the ISP as shown in Figure 13.2. This is done by masking client-​
specific information to avoid the compliance issues.
Data Set Information: The following attributes are available in a data set for
analysis.

• Network traffic classification data set

Variable Description

Flow ID Information about end-​to-​end network traffic generated


Source IP IP address of the source host
Source Port Network port number of sources
Destination IP IP address of the target host/​server
Destination Port Network port number of destinations
Protocol Network protocol number
Protocol Type Type of protocol: TCP /​UDP
L7 protocol OSI Layer seven protocol used in the connection
Application name Target application, which is accessed in network session

A total of 1048575 observations were made with nine network attributes.

• Network bandwidth data set

Variable Description

Timestamp Date and time of the details captured


Average TX Mbps Average transmit speed on the link on that day in Mbps
Average RX Mbps Average receive speed on the link on that day in Mbps
Classification of Network Traffic on ISP Link during COVID-19 217

Variable Description

Peak TX Mbps Peak transmit speed on the link on that day in Mbps
Peak RX Mbps Peak receive speed on the link on that day in Mbps
Max_​users Maximum number of user connections on that day
Bandwidth Bandwidth utilized on the link in a day

A total of 356 observations with 7 attributes including time series details.


Bandwidth details were captured for one year—​from July 17, 2019, to July 6,
2020, on a daily basis.

13.2.1 Network Topology
The network and application performance is measured through several
performance metrics such as bandwidth, throughput, disk time, number
of packets send/​Recv per sec, and number of bytes send/​Recv per second.
The network used in the experiment has several tools to collect network and
application metrics as in Figure 13.3. The metrics are collected on the hosts
and the communication link between the two end points.
The tables below show the performance metrics categorized by the
tool used:

Network monitor Performance monitor

Flow ID Bandwidth
Source IP Transfer rate
Destination IP Receive rate
Source port

FIGURE 13.3
Network performance metric tools.
218 Machine Learning Algorithms and Applications in Engineering

Network monitor Performance monitor

Destination port
Protocol
Protocol type
L7_​protocol
Application name
Max_​Users

13.2.2 Data Cleaning and Preprocessing


Preprocessing is a crucial step in transforming real-​world data into a usable
format. Certainly, real-​world data is frequently incomplete and noisy in spe-
cific behavior. To put it another way, most of the data needed to analyze using
data mining techniques comes from the real world and is incomplete and
inconsistent (containing errors, outlier values). As a result, pre-​processing
procedures are required before using data mining techniques to improve the
data quality, thereby contributing to improving the accuracy and efficiency of
the resultant data mining task. Due to the patterns of network traffic, which
have diverse types of format and dimensions, pre-​processing techniques
are necessary and important in network traffic analysis. All the categorical
variables are converted to numeric variables. All the categories are converted
into numeric levels except “Application Name,” which is the target variable
for classification.

13.2.3 Data Visualization with Tableau


It’s observed that Google, HTTP, and SSL connections are the top five
applications in the network traffic as shown in Figure 13.4. YouTube,
Windows updates, and Skype also are high contributors to the traffic flow on
the internet link.
From the above trend, it is shown that bandwidth requirements increased
almost exponentially from March 2020 onwards as shown in Figure 13.5. This
can be related with lockdown situations due to COVID-​19 where most of
the workforce of many organizations and institutions were forced to work
from home. This has increased the internet utilization across the globe. The
number of internet users increased drastically from 1 April 2020 onwards as
shown in Figure 13.6. As much as 7,000 to 12,000 users were connected in the
period of May to July 2020.
Peak transmit and receive speed is in line with the bandwidth usage as
observed for the one-​year period as shown in Figure 13.7. Transmit speed is
seen high increase during lockdown period.
newgenrtpdf
Classification of Network Traffic on ISP Link during COVID-19 219
Top five applications in the network traffic.
FIGURE 13.4
220 Machine Learning Algorithms and Applications in Engineering

FIGURE 13.5
Traffic increased almost exponentially from March 2020.

FIGURE 13.6
Increased internet users from April 1, 2020.

13.3 Traffic Classification Using Classification Algorithms


The process followed for network traffic classification model is as follows
(Figure 13.8):
Classification of Network Traffic on ISP Link during COVID-19 221

FIGURE 13.7
High increase during lockdown period.

FIGURE 13.8
Network traffic classification model.
222 Machine Learning Algorithms and Applications in Engineering

The cleaned data is set ready for classification as follows.

13.3.1 Comparison of Classifiers
The simulation tool provides precise results regarding the applied methods,
such as accuracy, training time, and recall, after various machine learning
algorithms have been implemented. The four different classifiers used in this
exercise are KNN, decision tree, random forest, and naïve Bayes.
When compared to other algorithms, the random forest algorithm produces
extremely accurate results. The accuracy results of applying these machine
learning techniques are compared in the chart in Figure 13.9.
Classification accuracy of 69 percent has been achieved using Naïve Bayes
algorithm. With k=​4, the classification accuracy of 89 percent is achieved
with KNN algorithm. Setting k=​3 and running the code again gives us better
results as shown below. Accuracy increased to 91 percent when the value
of k is reduced from 4 to 3. The decision tree gives improved accuracy of
92.89 percent compared to KNN. Random forest gives the highest accuracy
of 99.9 percent among all the algorithms attempted in classifying the network
traffic.

Classification Accuracy (%)


100%

50%

0%
KNN(k=4) KNN(k=3) Decision Tree Random Forest Naïve Bayes
FIGURE 13.9
Applying the accuracy results.
Classification of Network Traffic on ISP Link during COVID-19 223

13.4 Bandwidth Requirement Prediction Using


Time Series Model (ARIMA)
The time series data of bandwidth utilization for one of the ISP links is given
in the following table:

Average TX Average PX Peak TX Peak RX Bandwidth


TimeStamp (Mbps) (Mbps) (-​Mbps) (Mbps) Max Users (GB)

2019-​07-​17 71.75 117.1 172.6 207.39 6509 782.06


15:00:00
2019-​07-​18 47.06 106.67 90.46 204.74 6661 747.22
15:00:00
2019-​07-​19 44.98 36.31 135.53 153.86 6895 781.23
15:00:00
2019-​07-​20 27.96 44.11 45.54 94.64 6052 711.90
15:00:00
2019-​07-​21 17.48 30.51 36.36 48.91 5864 707.87
15:00:00
2019-​07-​22 36.77 81.78 64.52 163.03 6309 716.80
15:00:00
2019-​07-​23 41.96 97.95 77.84 173.13 6217 733.06
15:00:00

There are no null values as this is a TCP dump from the tool. All the network
parameters captured for the specified time stamp are given below.

13.4.1 Feature Extraction Selection


Before using data mining techniques, feature extraction selection is used as a
pre-​processing strategy. It is used to increase the performance of data mining
algorithms by removing redundant or irrelevant features. By selecting only
a subset of the original characteristics, feature extraction selection methods
generate a new set of attributes. Its primary purpose is to minimize the
224 Machine Learning Algorithms and Applications in Engineering

dimensionality of a data set in order to improve network traffic analysis. The


correlation matrix of the variables is obtained in a data set to identify the rele-
vant features for bandwidth prediction. From the correlation coefficients, the
following can be observed:

• High positive correlation between average TX speed and bandwidth


• High positive correlation between peak TX speed and bandwidth
• High positive correlation between maximum users and bandwidth
• Moderate relationship between RX speeds and bandwidth.

Hence transmit speed parameters and user count are used as features in the
prediction model.

• Average TX speed (Mbps)


• Peak TX speed (Mbps)
• Maximum users

The RX parameters was dropped from the data set for further modelling.
Then timestamp column was set as an index for the data set, which is required
for time series modelling.

The subplots of all the variables with respect to index (timestamp) provides
insights into data variation and trend as shown below in Figure 13.10.
There is substantial increase in network throughput post April 2020. This
is exactly when most of work started happening online across the industries.
This has put a heavy load on ISPs as almost all entire workforce across the
world is working remotely as offices were shut.

13.4.2 Auto-​Regressive Integrated Moving Average Model (ARIMA)


In order to predict the network traffic, the ARIMA model is used in conjunc-
tion with a time series model. To anticipate the bandwidth demand, we’ll
Classification of Network Traffic on ISP Link during COVID-19 225

FIGURE 13.10
Subplots of index of timestamp.

utilize the ARIMA module in Python to create a time series model. It’s critical
to examine the trends and seasonability of the time series data before moving
further with the TS model construction.
In comparison to 2019 and 2020, there is a significant difference in band-
width utilization as shown in Figure 13.11.

13.4.3 Time Series Data: Trends and Seasonality


Finding trends in a time series can be done in a variety of ways. A rolling
average is a common technique that includes taking the average of the points
on either side of each time point. It’s worth mentioning that the number of
points is influenced by the size of the window chosen. Noise and season-
ality are smoothed out as a result of utilizing the average. Considering a
window of 15 days, the trend line of bandwidth usage also shows a gradual
rise with time. As a result, when compared to the prior plots in Figure 13.11,
the majority of the seasonality has been erased as in Figure 13.12. The same
plots for TX speeds can be visualized in Figure 13.13 and number of users
connected to ISP link in Figure 13.14.
226 Machine Learning Algorithms and Applications in Engineering

FIGURE 13.11
Difference in bandwidth.

FIGURE 13.12
Removing the seasonality.

FIGURE 13.13
Visualize under TX speeds.
Classification of Network Traffic on ISP Link during COVID-19 227

FIGURE 13.14
Maximum number of users.

FIGURE 13.15
Upward trends of individual network from April 2020.

To compare the trends of individual network parameters, all the rolling


averages are plotted in a single graph. It is observed that there is an upward
trend in all the parameters from April 2020 onwards.

13.4.4 Time Series Data: Seasonal Patterns


To analyze seasonality more readily, remove the trend from a time series. This
is one approach to considering the seasonal elements of your day-​time TA’s
series. Subtract the original signal from the trend you calculated earlier to
get rid of the trend (rolling mean). However, the quantity of data points you
averaged will influence this data.
228 Machine Learning Algorithms and Applications in Engineering

FIGURE 13.16
First order difference of the ‘bandwidth’ data series.

FIGURE 13.17
First order differencing plots of TX speeds.

Another method for removing the trend is “differencing,” which involves


comparing the differences between successive data points as shown in
Figure 13.16 (known as “first-​order differencing” because you’re simply com-
paring one data point to the one before it).
Similarly, first-​order differencing plots of TX speeds shown in Figure 13.17
and number of users shown in Figure 13.18 show certain seasonability in
the data.
Classification of Network Traffic on ISP Link during COVID-19 229

FIGURE 13.18
First order differencing plots of maximum users.

FIGURE 13.19
First order differencing of all network parameters.

All the time series variables are plotted again on the same graph to see
how they look like. The first-​order differencing of all network parameters is
plotted in one graph as in Figure 13.19 to recheck the seasonal pattern if any.
From the above visualization, certain amount of seasonability in the time
series data for all the variables can be observed as in Figure 13.20 and also an
upward trend from the second quarter of 2020 onward.
230 Machine Learning Algorithms and Applications in Engineering

FIGURE 13.20
Seasonability in the time series data for all the variables.

13.5 Implementing the ARIMA Model


Splitting the data into train and test sets

Training data

Testing data
Classification of Network Traffic on ISP Link during COVID-19 231

13.5.1 Building ARIMA Model


The model building took 8.62 seconds.

For forecasting the next six values of bandwidth, we have six observations
in data set. We will then map these forecasts to start index of the test data for
comparison.
232 Machine Learning Algorithms and Applications in Engineering

13.5.2 Printing the Forecasted Values of Bandwidth

Let us concatenate the forecasted data frame with original data frame and
plot it (Figure 13.21).
Using the ARIMA model that was built, the prediction of the bandwidth
requirement of next two months is done and checked if the predictions are
in line with current spike in network utilization due to COVID situation. The
prediction of plot of this two-​month forecast along with original data set is
done to calculate the bandwidth required in August and September 2020 as
shown in Figure 13.22. It is also observed the range of forecast values show
the interval within which the predictions will fall for future dates. As we
can see, the bandwidth requirement will remain in the range of 800 to 1000
during August and September 2020 period as per the time series model as in
Figure 13.23.

FIGURE 13.21
The predictions are in line with high bandwidth utilization in 2020.
Classification of Network Traffic on ISP Link during COVID-19 233

FIGURE 13.22
Bandwidth required in August and September 2020.

FIGURE 13.23
Bandwidth range of 800–​1000 during August and September 2020.
234 Machine Learning Algorithms and Applications in Engineering

FIGURE 13.24
Evaluation of this proposed model using the diagnostic plot.

13.5.3 Model Evaluation
The evaluation of this proposed model can be done using the diagnostic plot
is shown in Figure 13.24.
Standardized Residuals: The residual errors appear to have a uniform vari-
ance and fluctuate around a mean of zero. The density plot indicates a normal
distribution with a mean of zero.
Normality: The dots on the Q–​Q plot should all be completely aligned with
the red line. Any large variances would indicate a skewed distribution.
The ACF plot, also known as the correlogram, illustrates that the residual
errors are not autocorrelated. Any autocorrelation would imply that the
residual errors have a pattern that isn’t explained by the model. As a result,
you’ll need to add more Xs (predictors) to the model. In general, it appears
to be a decent match.

13.6 Conclusion: Business Benefits of the Models


13.6.1 Network Traffic Classification
Network traffic categorization is a method that divides computer network
traffic into different classifications based on various criteria (such as port
Classification of Network Traffic on ISP Link during COVID-19 235

number or protocol). Using classified network traffic, you may do tasks like
monitoring, discovery, control, and optimization. The overall purpose of net-
work traffic classification is to improve network performance.
The packets can be labelled once they have been classified as belonging to a
specific application. These markings or flags will assist the network router in
determining the proper service policies for certain network flows.
This study’s classification approach can be used in conjunction with net-
work or packet monitoring software. This allows us to create a real-​time
dashboard that shows application-​specific network use in real time. This
will aid in the rapid and efficient regulation of traffic flow across any
internet link.

13.6.2 Time Series Prediction


Due to the introduction of new technology, industries, and applications,
internet traffic has increased dramatically during the last decade. The
Internet of Things (IoT), Cloud computing, and data-​center applications
are all growing in popularity. Networks must cope with ever-​increasing
traffic demands while maintaining a high level of service to users. As a
result, effective networking equipment usage has become critical. If net-
work traffic can be properly forecast, this task can be completed effectively.
Many networking sectors, such as energy conservation, network resource
management, and wireless sensor networks, can benefit from accurate traffic
forecast.
One advantage of traffic prediction is that it saves energy. Traffic pre-
diction can be utilized in internet core network routers/​switches to save
a large amount of power. The number and sophistication of processors
employed in these network devices are expanding in response to rising
traffic demands and computing requirements, resulting in increased power
consumption. High equipment power consumption and accompanying
increases in cooling expenses result in increased network operational costs.
If network traffic can be precisely forecast, additional processors in these
core devices can be turned down to save power during low traffic times.
Construction of greener traffic-​aware networks can accurately estimate on
future traffic loads.
Second, in order to make an optimum use of network resources, traffic fore-
cast is essential. Accurate network traffic prediction at access points provides
efficient resource allocation to ensure acceptable service quality.
Finally, when new applications such as e-​commerce and online banking
emerge, identifying and preventing network abuse is becoming increasingly
challenging due to increased traffic and network complexity. Anomaly detec-
tion is one method of preventing network abuse. An assault can be detected
by a considerable divergence from typical traffic behavior. The accuracy
of traffic forecast is directly proportional to the performance of anomaly
detection.
236 Machine Learning Algorithms and Applications in Engineering

Bibliography
[1] Shahraki, Amin, Mahmoud Abbasi, Amir Taherkordi, and Anca Delia Jurcut.
“Active learning for network traffic classification: A technical survey.” arXiv pre-
print arXiv:2106.06933 (2021).
[2] Callegari, Christian, Pietro Ducange, Michela Fazzolari, and Massimo Vecchio.
“Explainable Internet Traffic Classification.” Applied Sciences 11, no. 10
(2021): 4697.
[3] Abbasi, Mahmoud, Amin Shahraki, and Amir Taherkordi. “Deep learning
for network traffic monitoring and analysis (NTMA): A survey.” Computer
Communications 170 (2021): 19–​41.
[4] Dong, Shi. “Multi class SVM algorithm with active learning for network traffic
classification.” Expert Systems with Applications 176 (2021): 114885.
[5] Zhongsheng, Wang, Wang Jianguo, Yang Sen, and Gao Jiaqiong. “Traffic identi-
fication and traffic analysis based on support vector machine.” Concurrency and
Computation: Practice and Experience 32, no. 2 (2020): e5292.
[6] Salman, Ola, Imad H. Elhajj, Ayman Kayssi, and Ali Chehab. “A review on
machine learning-​based approaches for internet traffic classification.” Annals of
Telecommunications 75, no. 11 (2020): 673–​710.
[7] Network Based Application Recognition (NBAR)—​ Cisco, 2009: www.cisco.
com/​c/​en/​us/​produ​cts/​ios-​nx-​os-​softw​are/​netw​ork-​based-​appl​icat​ion-​reco​
gnit​ion-​nbar/​index.html. (Accessed March 12, 2021)
[8] Palo Alto Networks, Next-​Generation Firewall Overview, 2011.
[9] www.paloa​lton​etwo​rks.com/​cont​ent/​dam/​paloa​lton​etwo​rks-​com/​en_​US/​
ass​ets/​pdf/​dat​ashe​ets/​firew​all-​featu​res-​overv​iew/​firew​all-​featu​res-​overv​iew.
pdf. (Accessed March 15, 2021)
[10] Mellia, Marco, Antonio Pescapè, and Luca Salgarelli. “Traffic classification
and its application to modern networks.” Computer Networks (1999) 53, no. 6
(2009).
[11] Risso, Fulvio, Mario Baldi, Olivier Morandi, Andrea Baldini, and Pere Monclus.
“Lightweight, payload-​ based traffic classification: An experimental evalu-
ation.” In 2008 IEEE International Conference on Communications, pp. 5869–​5875.
IEEE, 2008.
[12] Avalle, Matteo, Fulvio Risso, and Riccardo Sisto. “Efficient multistriding of large
non-​deterministic finite state automata for deep packet inspection.” In 2012 IEEE
International Conference on Communications (ICC), pp. 1079–​1084. IEEE, 2012.
[13] Crotti, Manuel, Maurizio Dusi, Francesco Gringoli, and Luca Salgarelli. “Traffic
classification through simple statistical fingerprinting.” ACM SIGCOMM
Computer Communication Review 37, no. 1 (2007): 5–​16.
[14] Bujlow, Tomasz, Tahir Riaz, and Jens Myrup Pedersen. “A method for classifi-
cation of network traffic based on C5. 0 Machine Learning Algorithm.” In 2012
International Conference on Computing, Networking and Communications (ICNC),
pp. 237–​241. IEEE, 2012.
[15] Nguyen, Thuy T.T., and Grenville Armitage. “A survey of techniques for internet
traffic classification using machine learning.” IEEE Communications Surveys &
Tutorials 10, no. 4 (2008): 56–​76.
Classification of Network Traffic on ISP Link during COVID-19 237

[16] Ubik, Sven, and Petr Žejdl. “Evaluating application-​layer classification using a
machine learning technique over different high-​speed networks.” In 2010 Fifth
International Conference on Systems and Networks Communications, pp. 387–​391.
IEEE, 2010.
[17] Bujlow, Tomasz, and Jens Myrup Pedersen. “Multilevel classification and
accounting of traffic in computer networks.” Classification and Analysis of
Computer Network Traffic 167.
[18] Bujlow, Tomasz, and Jens Myrup Pedersen. “A practical method for multi-
level classification and accounting of traffic in computer networks” Aalborg
Universitet (2014).
[19] Jun, Li, Zhang Shunyi, Lu Yanqing, and Zhang Zailong. “Internet traffic clas-
sification using machine learning.” In 2007 Second International Conference on
Communications and Networking in China, pp. 239–​243. IEEE, 2007.
[20] Bujlow, Tomasz, and Valentin Carela-​Espanol. “Comparison of deep packet
inspection (DPI) tools for traffic classification” (UPC-​DAC-​RR-​CBA-​2013-​3 ed.)
Universitat Politècnica de Catalunya. www.ac.upc.edu/​app/​resea​rch-​repo​rts/​
html/​resear​ch_​c​ente​r_​in​dex-​CBA-​2013,en.html (2013).
[21] Dusi, Maurizio, Francesco Gringoli, and Luca Salgarelli. “Quantifying the
accuracy of the ground truth associated with Internet traffic traces.” Computer
Networks 55, no. 5 (2011): 1158–​1167.
[22] Cascarano, Niccolò, Luigi Ciminiera, and Fulvio Risso. “Optimizing deep
packet inspection for high-​speed traffic analysis.” Journal of Network and Systems
Management 19, no. 1 (2011): 7–​31.
[23] Alcock, Shane, and Richard Nelson. “Measuring the accuracy of open-​source
payload-​based traffic classifiers using popular internet applications.” In 38th
Annual IEEE Conference on Local Computer Networks-​ Workshops, pp. 956–​
963.
IEEE, 2013.
[24] Goss, Ryan, and Reinhardt Botha. “Deep packet inspection—​ Fear of the
unknown.” In 2010 Information Security for South Africa, pp. 1–​5. IEEE, 2010.
[25] Valenti, Silvio, Dario Rossi, Alberto Dainotti, Antonio Pescapè, Alessandro
Finamore, and Marco Mellia. “Reviewing traffic classification.” In Data Traffic
Monitoring and Analysis, pp. 123–​147. Springer, Berlin, Heidelberg, 2013.
[26] Wei, Yong, Z. Yun-​Feng, and Li-​chao Guo. “Analysis of message identification
for OpenDPI.” Computer Engineering (2011): S1.
[27] Alcock, Shane, and Richard Nelson. Libprotoident: Traffic Classification Using
Lightweight Packet Inspection. Technical report, University of Waikato, 2012.
[28] Levandoski, Justin. “Application layer packet classifier for Linux.” http://​l7-​fil​
ter. sourceforge. net/​(2008).
[29] Dainotti, Alberto, Walter De Donato, and Antonio Pescapé. “Tie: A community-​
oriented traffic classification platform.” In International Workshop on Traffic
Monitoring and Analysis, pp. 64–​74. Springer, Berlin, Heidelberg, 2009.
[30] Khakpour, Amir R., and Alex X. Liu. “High-​speed flow nature identification.”
In 2009 29th IEEE International Conference on Distributed Computing Systems,
pp. 510–​517. IEEE, 2009.
[31] Kim, Hyunchul, Kimberly C. Claffy, Marina Fomenkov, Dhiman Barman,
Michalis Faloutsos, and KiYoung Lee. “Internet traffic classification demys-
tified: myths, caveats, and the best practices.” In Proceedings of the 2008 ACM
CoNEXT conference, pp. 1–​12. 2008.
238 Machine Learning Algorithms and Applications in Engineering

[32] Aceto, Giuseppe, Alberto Dainotti, Walter De Donato, and Antonio Pescapé.
“PortLoad: taking the best of two worlds in traffic classification.” In 2010
INFOCOM IEEE Conference on Computer Communications Workshops, pp. 1–​5.
IEEE, 2010.
[33] Kumar, Sailesh, Sarang Dharmapurikar, Fang Yu, Patrick Crowley, and Jonathan
Turner. “Algorithms to accelerate multiple regular expressions matching for
deep packet inspection.” ACM SIGCOMM Computer Communication Review 36,
no. 4 (2006): 339–​350.
[34] Cascarano, Niccolo, Pierluigi Rolando, Fulvio Risso, and Riccardo Sisto.
“iNFAnt: NFA pattern matching on GPGPU devices.” ACM SIGCOMM Computer
Communication Review 40, no. 5 (2010): 20–​26.
[35] Bujlow, Tomasz, Valentín Carela-​Español, and Pere Barlet-​Ros. “Extended inde-
pendent comparison of popular deep packet inspection (DPI) tools for traffic
classification.” (2014). Universitat Politècnica de Catalunya. www.ac.upc.edu/​
app/​resea​rch-​repo​rts/​html/​resear​ch_​c​ente​r_​in​dex-​CBA-​2014,en.html
[36] Fusco, Francesco, and Luca Deri. “High speed network traffic analysis with
commodity multi-​core systems.” In Proceedings of the 10th ACM SIGCOMM
Conference on Internet Measurement, pp. 218–​224. 2010.
newgenprepdf

239

14
Integration of AI/​Ml in 5G Technology
toward Intelligent Connectivity,
Security, and Challenges

Devasis Pradhan,1,* Prasanna Kumar Sahu,2 Rajeswari,1 Hla Myo Tun,3


and Naw Khu Say Wah3
1
Department of Electronics and Communication Engineering, Acharya Institute of
Technology, Bangalore
2
Department of Electrical Engineering, National Institute of Technology, Rourkela
Department of Electronic Engineering, Yangon Technological University, Yangon,
3

Myanmar
[email protected]
*

CONTENTS
14.1 Introduction.............................................................................................. 240
14.2 Overview of 5G........................................................................................ 240
14.2.1 Enabling Technologies............................................................... 241
14.2.2 Key Enablers for 5G................................................................... 241
14.2.3 5G Requirements........................................................................ 243
14.3 Artificial Intelligence (AI) and Machine Learning.............................. 243
14.3.1 Amalgam of AI and ML in 5G.................................................. 244
14.3.2 Security and Privacy.................................................................. 244
14.3.3 AI Learning-Based Scheme for 5G........................................... 246
14.4 Wireless AI................................................................................................ 246
14.5 AI in Context with 5G............................................................................. 248
14.6 Intelligent Connectivity........................................................................... 249
14.7 5G Security................................................................................................ 250
14.8 Challenges................................................................................................. 250
14.8.1 5G Complexity............................................................................ 250
14.8.2 Security and Privacy.................................................................. 251
14.8.3 Trade-off in Speed...................................................................... 251
14.9 Research Scope......................................................................................... 251
14.10 Conclusion................................................................................................ 252

DOI: 10.1201/9781003104858-14 239


240

240 Machine Learning Algorithms and Applications in Engineering

14.1 Introduction
In the ceaselessly developing correspondence organization engineering, to
coordinate different scopes of gadgets with remarkable prerequisites for
an alternate organization, parameters have brought about refined difficul-
ties for network security. The new improvements in 5G networks and past
are working with the vivid development of information correspondence
by giving higher information rates. 5G is a versatile network confronting
many difficulties to meet the remarkable developing requests for admit-
tance to remote administrations with super low idleness and, further, high
information rates. 5G today is the central innovation of many state-​of-​the-​art
advancements like the IoT, smart grid, automated aeronautical frameworks,
and self-​driving vehicles. 5G is needed to be described by high adaptability in
plan and assets by the executives and allotted to fulfill the expanding needs
of this heterogeneous network and clients or end-​users[1–​3].
Numerous wireless examinations driving gatherings foresee that artifi-
cial intelligence (AI) as the next huge “game-​evolving” innovation, ready
to furnish 5G with the adaptability and the knowledge required. Hence
numerous analysts have examined the proficiency of this hypothesis in
numerous parts of 5G remote interchanges including regulation, channel
coding, obstruction of the executives, planning, 5G slicing, reserving, energy
productivity, and digital protection. ML and AI algorithms can be utilized
to process and dissect cross-​space information that would be needed in 5G
in a considerably more effective manner empowering speedy choices and
as such facilitating the organization’s intricacy and decreasing the support
cost. The cross-​space information incorporates geographic data, designing
boundaries, and different information to be utilized by AI and ML to all the
more likely gauge the pinnacle traffic, streamline the organization for limit
extension, and empower more canny inclusion through powerful imped-
ance estimations [4].

14.2 Overview of 5G
The quick reception of 5G innovation is promising a stunning number of new
gadgets. For instance, the Cisco Annual Internet Report (2018–​2023) estimates
that “Machine-​to-​Machine (M2M) communication will develop 2.4 overlays,
from 6.1 billion out of 2018 to 14.7 billion by 2023. There will be 1.8 M2M com-
munication for every individual from the worldwide populace by 2023.” The
outstanding development in associated gadgets alongside the presentation of
5G innovation is relied upon to cause a test for productive and dependable
organization’s asset assignment. In addition, the gigantic organization of the
241

Integration of AI/Ml in 5G Technology 241

Internet of Things and associated gadgets to the internet might pose a genuine
danger to the organization’s security assuming they are not taken care of appro-
priately [5,6]. The 5G networks are relied upon to help a lot higher level het-
erogeneity (as far as associated gadgets and networks) when contrasted with
their archetypes. For example, 5G supports intelligent vehicles, shrewd homes,
savvy structures, and shrewd urban areas. Also, the IoT in the 5G design will
include more strong and versatile strategies to deal with the basic security issues
both at the organization and gadget sides. The security of such a network will
be considerably more convoluted due to the external interruption just as the
nearby interruption. Artificial intelligence and ML can give arrangements by
grouping delicate security connections in the middle, for example, character,
validation, and confirmation. The security and security in 5G-​IoT will cover
every one of the layers like character assurance, security, and protecting end-​
to-​end users. For example, the key validation system from end-​gadget to core
network and onward to the specialist co-​op, while hiding the key identifier, is
as yet a mind-​boggling issue. AI and ML can likewise assume a significant part
in key validation alongside adequately limiting the disguising attacks.

14.2.1 Enabling Technologies
5G will empower its availability billions of gadgets in the IoT and the
associated world. Thus, there are three significant classifications of utiliza-
tion cases for 5G.

(a) Enormous machine to machine communication: It’s additionally


considered IoT, which includes associating billions of gadgets without
human intercession at a scale not seen previously. This has cutting-​
edge modern interaction and applications including agribusiness,
producing, and business correspondence.
(b) Low latency with ultra-​reliability: It incorporates ongoing control of
gadgets, modern advanced mechanics, vehicle-​to-​vehicle correspond-
ence likewise opening up another reality where far-​off clinical consid-
eration, strategy, and treatment are conceivable.
(c) Improved versatile broadband: It gives basically quicker information
speed and more prominent limit keeping the world associated. The
clever application will incorporate fixed remote Web access for homes
and outside broadcast application networks for individuals moving.

14.2.2 Key Enablers for 5G


The key enabling techniques have been elaborated in Table 14.1 with their
applicability. These key enablers help the 5G network to make more cogni-
tive and versatile application-​oriented networks.
newgenrtpdf
242
242
TABLE 14.1
Key Enablers for 5G Network

S. No. Techniques Applicability Support to Network Challenges

1 NFV (network function Asset usage further developed Data-​centric Disengagement, asset designation, decency,
virtualization) throughput, energy investment funds, communication income/​value improvement, versatility the

Machine Learning Algorithms and Applications in Engineering


decreased CAPEX/​OPEX, upgraded (everything as executives
QoE, more straightforward assistance)
movement /​support.
2 C-​RAN (Cloud-​RAN) Adaptability, energy /​power investment Administration situated BBU managements (e.g., participation,
funds, expanded throughput, communication with interconnection, bunching), energy-​
decreased deferral, flexibility to heterogeneous network mindful planning, fronthaul-​mindful asset
dynamic traffic, diminished CAPEX/​ (HetNet) designation
OPEX, simpler management of used
network
3 HetNet (Heterogeneous Expanded throughput, range use, energy Small cell network, D2D/​ Mitigating the interference versatile power
Network) productivity, inclusion extension M2M communication, control, dynamic mode choice and
Internet-​of-​Things offloading to underlay organization, gadget
revelation, bound together MAC plan
4 FDC (Full Duplex Range productivity, decreased inertness, Small cell network, D2D/​ SI decrease, cross-​layer asset the executives,
Communication) energy effectiveness M2M, intellectual radio power portion, impedance the board,
networks, multihop synchronization and time change in
transferring accordance with setting up FD transmission,
dynamic mode determination, planning a
MAC convention
5 RF-​EH (RF Energy Green communication with energy Small cell network, D2D/​ Muti-​client planning, progressed channel
Harvesting) efficiency M2M communication, securing, energy beamforming, gather/​
communicate time variation, obstruction
the board, SWIPT-​empowered asset
distribution
243

Integration of AI/Ml in 5G Technology 243

TABLE 14.2
5G Requirements as Compared to 4G

S. No. Verticals Remarks

1 Speed 1–​10 Gbps associations with end focuses in the field (for
example, not hypothetical greatest)
2 Latency 1 millisecond start to finish full circle delay—​dormancy
3 BW utilization 1000× transfer speed per unit region
4 Device to be connected 10–​100× number of associated devices
5 Accessibility Perception of 99.999% accessibility
6 Coverage 100%
7 Usage of energy 90% decrease in network energy use
8 Battery life of device As long as ten-​year battery life for low power, machine-​
type gadgets

14.2.3 5G Requirements
Lately, there have been a few perspectives about a definitive structure that
5G technology should take. There have been two perspectives on what 5G
ought to be:

(a) Hyper-​connectivity: This perspective on the prerequisites for 5G remote


frameworks means to take the current innovations including 2G, 3G,
4G, Wi-​Fi, and other significant remote frameworks to give higher
inclusion and accessibility, alongside more thick organizations. Aside
from having necessities to offer conventional types of assistance, a key
differentiation is empowered new administrations like machine to
machine (M2M) applications alongside extra Internet of Things (IoT)
applications [7]‌. This arrangement of 5G necessities could require
another radio innovation to empower low power, low throughput
field gadgets with long battery lifetimes of ten years or more.
(b) Radio-​Access Technology (RAT): This perspective on the 5G prerequisites
takes a more innovation-​driven view and sets details for information
rates, inertness, and other key boundaries. These necessities for 5G
would empower an unmistakable boundary to be made between 4G
or different administrations and the new 5G remote framework [8, 9].

14.3 Artificial Intelligence (AI) and Machine Learning


The idea of utilizing AI and ML in security isn’t new, yet their attainability
and execution prevalence generated interest with the advancement of deep
learning (DL) calculations. A large portion of the strategies previously aimed
244

244 Machine Learning Algorithms and Applications in Engineering

at the improvement of DL was devoted to displaying the assault designs


with specific attributes that are not vigorous in nature, yet with AI and ML,
it is normal that frameworks will turn out to be stronger toward new refined
dangers and assaults with dynamic attributes. This is because aggressors
utilize complex methods like confusion, polymorphism, or pantomime to
stay away from recognition. From bundle catching and examination to enor-
mous information bits of knowledge, AI and ML can be utilized to advise
the dangers not distinguished by ordinary methods. The example-​based
learning at the center is upheld by softwarization furthermore, and virtual-
ization gives spryness and vigor to convenient counter cyber attacks [10–​12].
Artificial intelligence calculations are being used for security and protec-
tion issues. The data security industry is creating an ever-​increasing number
of information that opens them to propel threats and AI could be an amazing
antitoxin. The original AI arrangements are zeroing in on examining infor-
mation, recognizing the threat, and helping people in the remediation plan.
The second era of AI will make the frameworks more independent and just
leave the basic help issues to people [10, 11].

14.3.1 Amalgam of AI and ML in 5G
An expanded transfer speed, higher range use, and further, high information
rates in 5G organizations have additionally enlarged the threat and security
scene from individual gadgets to the specialist co-​op network. Consequently,
the network ought to be savvy enough to manage these difficulties continu-
ously and ML and AI methods could assist with demonstrating these strong
unique calculations in recognizing network issues and furnish with the con-
ceivable arrangement progressively. In short to medium-​term plans, AI and
ML can be utilized to distinguish the dangers and counter them with hearty
and versatile security calculations [13]. Though, in the long haul, a com-
pletely mechanized security instrument is imagined for opportune reaction
to threats and cyberattacks.
The 5G network relied upon to help a lot higher level heterogeneity (as far
as associated gadgets and networks) when contrasted with its archetypes.
For example, 5G supports brilliant vehicles, shrewd homes, savvy structures,
and smart urban areas. Additionally, the IoT in 5G network design will
include more hearty and versatile strategies to deal with the basic security
issues both at the network and gadget sides [13, 14].

14.3.2 Security and Privacy


The security of such networks will be significantly more confounded as a
result of the external interruption just as the nearby interruption. AI and
ML can give arrangements by arranging delicate security connections in the
245

Integration of AI/Ml in 5G Technology 245

middle, for example, identity, confirmation, and protecting the user details.
The security and privacy in 5G-​IoT will cover every one of the layers like
personality insurance, protection, and E2E protection. Catering for security
and protection of information from these various frameworks with extraor-
dinarily unique security prerequisites becomes a monotonous undertaking
[15]. AI and ML with an outline of SBA and security prerequisites for dis-
tinctive end-​frameworks can identify and redress these issues progressively
by arranging and bunching surprising dangers. AI and ML can help in cre-
ating security instruments by making trust models, gadget security, and
information confirmation to give precise security for the entire 5G-​IoT net-
work [16,17].

TABLE 14.3
AI Techniques for 5G

AI Learning Leaning or
S. No. Scheme Training Model Application

1 Supervised ML and statistical Dynamic recurrence and transfer speed


logistic portion in self-​coordinated LTE dense little
cell arrangements. Path loss
SVM (Support Vector Prediction of Pathloss estimation
Machine)
Neural Network Channel learning to derive undetectable
Based channel state information (CSI) from a
recognizable channel
ANN and MLPs Demonstrating and approximations of true
capacities for interface spending plan and
proliferation misfortune for cutting edge
wireless network.
2 Unsupervised Gaussian Mixture Cooperative spectrum sensing for vehicular
Model (GMM) networks
Hierarchical Fault detection
clustering (HCM)
Affinity clustering Resource management for ultra-​small cell
model (ACM) network
3 Reinforcement Long short-​term Proactive asset allocation in LTE-​U
memory model Networks, formed as a non-​cooperative
game, which empowers SBSs to realize
which unlicensed channel, given the
drawn-​out WLAN movement in the
channels and LTE-​U traffic loads
Grad follower (GF) Empower femto-​cells (FCs) to independently
and MBM and astutely detect the radio climate
and tune their boundaries in HetNets, to
decrease intra/​inter-​level impedance
246

246 Machine Learning Algorithms and Applications in Engineering

14.3.3 AI Learning-​Based Scheme for 5G


(a) Supervised learning scheme: In supervised learning, each prepar-
ation model must be taken care of alongside its individual name.
The thought is that preparation of a learning model is an example
of the issue occasions with known optima, and afterward, utilize the
model to perceive ideal answers for new cases. An average under-
taking on directed learning is to foresee an objective numeric worth,
given a bunch of highlights, called indicators. This depiction of the
undertaking is called relapse. Move learning is a famous procedure
frequently used to group vectors. Basically, one would prepare a con-
volutional neural network (CNN) on an extremely huge data set, for
instance, Image Net [14], and afterward adjust the CNN on an alter-
nate vector data set. The great part here is preparing on the huge data
set is as of now done by certain individuals who offer the learned
loads for public examination use.[16–​18]
(b) Unsupervised Learning Scheme: In unsupervised learning, the prep-
aration information is unlabeled, and the framework endeavors
to learn with no direction. This strategy is especially helpful when
we need to distinguish gatherings of comparative attributes. Never
do we advise the calculation to attempt to recognize gatherings of
related characteristics; the calculation tackles this association without
intercession [16,17].
(c) Reinforcement Learning Scheme: It depends on a learning framework
frequently called a specialist, which responds to the climate. The spe-
cialist performs activities and gets prizes or punishments (negative
prizes) as a trade-​off for its activities. That implies that the specialist
needs to learn without help from anyone else making an approach
that characterizes the activity that the specialist ought to pick in a spe-
cific circumstance. The point of the support learning task is to expand
the previously mentioned award over the long run [16,17].

14.4 Wireless AI
5G communication and network management is enabled by machine
learning and deep learning. For every model, the benefits and weaknesses of
AI-​empowered procedures is discussed as follows:

(a) M-​MIMO and beamforming: Massive MIMO is one element of 5G. Using
countless radiating antennas, 5G can concentrate the transmission and
gathering of signal power into small regions. Be that as it may, a few
issues are identified with this innovation. AI/​ML has been applied in
247

Integration of AI/Ml in 5G Technology 247

massive MIMO to beat these issues. For example, a precise gauge of


the channel with basic assessment strategies and a sensible number
of pilots is tried in huge MIMO: the low intricacy least-​squares (LS)
assessor doesn’t accomplish an acceptable execution, while least mean
square error (MMSE) channel assessment is extremely intricate. AI and
deep learning have been likewise examined in enhancing the loads of
receiving wire components in enormous MIMO. Profound learning
and AI can foresee the client appropriation and in like manner improve
loads of antenna components, which can work on the inclusion in a
multi-​cell situation [5, 19].
(b) Channel coding: An observable component of the air interface of the 5G
is the utilization of new channel coding methods: Data channels utilize
low-​thickness equality check (LDPC) codes, and control channels
utilize polar codes [8]‌ . For example, polar codes can accomplish
astounding execution; however, it takes a few emphases to accom-
plish this exhibition, and it is absolutely impossible to anticipate how
quickly polar codes can arrive at this ideal presentation. Profound
learning is notable for its high parallelism structure, which can carry
out a single shot coding/​ deciphering. Subsequently, numerous
specialists foresee that profound learning-​based channel coding is a
hopeful technique to empower 5G NR. The DL-​based channel coding
can accomplish a decent scope of execution intricacy compromises, in
case the preparation is performed effectively as the decision of code-​
word length causes over-​fitting and under-​fitting [8, 9, 20].
(c) Energy-​efficient network: The usage of ICT is answerable for 2–​10 per-
cent of the world’s energy utilization in 2007, and it is relied upon to
for further development [21]. Additionally, over 80 percent of ICT is
from a radio access network (RAN), which is sent to meet the pin-
nacle traffic burden and stays on it regardless of whether the heap is
light. This goal can be accomplished by lessening the power utiliza-
tion of the base stations and cell phones. AI/​ML adapting along these
lines can help in building clever remote organizations that proactively
foresee the traffic and versatility of clients and conveyance benefits
just when mentioned, therefore, decreasing the power utilization in
radio access networks (RAN).[7–​9, 20]
(d) 5G network slicing: Two conspicuous elements of 5G are network sli-
cing and caching. The first permits administrators to convey diverse
assistance types in more than one network framework. The last option
predicts the substance that clients might demand productive use of
the capacity of the base station. 5G asset provisioning and storing
utilize the hypothesis of AI/​ML [22].
(e) Cognitive radio: Radio access is scant, and there is an expanding
interest in remote traffic. Insightful remote organization executives is
the way forward to meet these expanding requests. AI/​ML can be a
248

248 Machine Learning Algorithms and Applications in Engineering

promising to be included for asset assignment in 5G networks. DL can


be a decent option for the avoidance of interference in the network,
usage of spectrum, multi-​path utilization with less path loss, interface
variation, multi-​channel access, and traffic blockage [23].
(f) Modulation: AMC is a central strategy in non-​cooperative communica-
tion frameworks. Modulation regulation is one errand that can help in
arranging the balance sort of a got signal, which is an important stage
toward comprehension and detecting the remote climate. Great detecting
and variation work on ghastly proficiency and obstruction alleviation.
Profound learning-​based AMC frameworks comprise three parts: The
initial segment is signal handling to upgrade the nature of the got tests,
recurrence offset rectification, gain control, enhancers, and separating.
The subsequent part includes the extraction of elements like the abun-
dance, stage, and recurrence of the got signal. The last part is a sign clas-
sifier: characterization of the balance types. AI/​ML can accomplish the
high exactness of adjustment characterization [23,24].

14.5 AI in Context with 5G


With regards to 5G and the improvement of versatile organizations, AI and
ML are reciprocally utilized, yet they contrast from one another. AI is an
expansive idea that does specific undertakings astutely and nearer to people.
It depends on ML to gather the information and investigate the example from
which the product framework learns and improves, and this makes machines
more intelligent. On the opposite side, ML is a subset of AI and is viewed as
a use of AI to permit machines to access information and assess on their own.
ML is otherwise called a present status of the craft of AI [25]. The complexity
of 5G can be overcome using two fundamental approaches:

(a) Basic: The fundamental methodology where AI/​ ML is utilized to


play out some essential assignments is dependent on some preset
calculations or algorithms.
(b) Advanced: In this developed methodology, AI/​ML is utilized to be more
mindful of settings and gain from the encompassing circumstances
and acts as:
(i) This methodology has arisen with the prevalence of the internet
and the enormous measure of the created data.
(ii) Instead of training the PC to do everything, it is smarter to code
them to think like people and give them admittance to the tre-
mendous data empowered by the Internet and 5G.
249

Integration of AI/Ml in 5G Technology 249

14.6 Intelligent Connectivity
It is a new idea, which is based on a combination of three significant
innovations, 5G and IoT, which is intended to fill in as a way to speed
up the advancement of troublesome computerized administrations. This
new idea works with the association of gadgets through a quick and low
inactivity portable organization, that is 5G, which gathers advanced data
through the machines and sensors, which is the capacity of IoT. At that
point, examinations and contextualizations by AI/​ML lastly creates sig-
nificant results that are helpful for the clients [25,26]. This would empower
new groundbreaking capacities in the greater part of the business areas, for
example transport, producing, medical care, public well-​being, security, and
so forth.

(a) Planning of network: The operator will utilize AI to further develop


the organizational scope quantification, which will prompt a decrease
in expenses and better execution of the organization. Artificial intel-
ligence and ML can be applied to foresee and conjecture traffic by
distinguishing traffic designs and as such learning on the web and
assisting with the mechanization of choices. This voids the require-
ment for over-​arrangement just like the case with the conventional
organization limit arranging [27].
(b) Performance of network to be monitored: This permits a network regulator
to gain as a matter of fact while it upgrades the network.
(c) Customer satisfaction: AI will assist with improving and dealing with
the client experience by utilizing the IoT information that uncovers
significant shopper bits of knowledge with regards to real-​ time
circumstances. This aids the purchasers and furnishes them with an
encounter tailor-​made to their life [27,28].
(d) Product life cycle: AI/​ ML assists with overseeing and further
developing the item life cycle by utilizing the information that
portrays the current and recorded item experiences, main drivers
and relationships, future results, and suggested upgrades. This
assists with changing the item life cycle from an information in the
executives instrument to an insightful dynamic framework toward
cognitive connectivity [28,29]
(e) Revenue generation: The integration of AI/​ML in 5G is as of now
occurring and will lessen the capital consumption and as such further
develop the income stream as well as to utilize the devices more effect-
ively with less utilization of energy [29,30].
250

250 Machine Learning Algorithms and Applications in Engineering

14.7 5G Security
Generally, AI and ML calculations are information hungry in nature, which
implies that information is expected to prepare the model for powerful
working. In the period of 5G, information age, stockpiling, and the man-
agement of data are quite easy as we have high computational power, out-
standing information development, and information sources. The network
can be kept up with, accessed, and examined for potential threats, attacks,
and weaknesses utilizing AI and ML at a lower cost of figuring and a reason-
able framework [21, 30–​32].
AI and ML models can be utilized to recognize dubious exercises con-
tinuously by dissecting the designs and boundaries of network movement.
Order calculations can be utilized to distinguish inconsistencies by
observing organization boundaries such as throughput and organization
blunder logs. Grouping algorithms can be utilized to classify different
sorts of dangers and also escape clauses in network security. The models,
for example, measure induction assaults and generative adversarial
networks (GAN) can create counterfeit datasets and these are new safety
efforts just like testing and carrying out advanced security conventions
and calculations. The exploration in creating private AI and ML models
have seen some huge improvement in secure computation, encryption,
protection, and unified learning. Half and half models are made by taking
on methods from various fields to make models effective, quicker, and
summed up [21, 33, 34].

14.8 Challenges
14.8.1 5G Complexity
AI/​ML algorithms should be deployed at the time of proper communica-
tion in wireless gadgets. Be that as it may, numerous remote gadgets have
restricted memory and registering capacities, which isn’t appropriate for
complex calculations. The assortment of huge examples and preparing pro-
found learning models take extensive time, which is a huge hindrance to
convey them on some remote gadgets having restricted power and cap-
acity. Now and again, the higher the quantity of tests and the more huge
the preparation time are, the higher the precision of acknowledgment of the
sign and organization highlights is. Obtaining more examples and preparing
the models for longer occasions bring about lethargic input. Thus, machine
learning models ought to be intended to accomplish best precision with less
examples and within a brief time frame [34].
251

Integration of AI/Ml in 5G Technology 251

14.8.2 Security and Privacy


Safeguarding the security of the clients is the essential worry of portable and
specialist co-​ops. One of the principal challenges in remote AI is the means
by which one can empower the preparation on a data set having a place with
clients without sharing the information and putting the individual data of
clients in danger. It is important to have a security way to deal with the mix of
profound learning in remote interchanges [35]. The security of AI/​ML models
itself is another test, as the neural network is pruned to antagonistic attacks.
Attackers can influence the preparation interaction by infusing counterfeit data
sets; such infusion can bring down the precision of the models and yield wrong
plan, which may influence the execution in an organization. The examination of
ML or AI with regard to security, by and large, remains shallow [35,36].

14.8.3 Trade-​off in Speed
The dependability of these procedures is definitely not exactly conven-
tional strategies in remote interchanges in taking care of certain issues. For
example, profound learning can contend with LS and MMSE in remote
divert assessment in massive MIMO; however, lethargic input portrays these
methods. AI/​ML induction might lengthen the framework reaction time.
This is because not most remote gadgets approach cloud computing, and
regardless of the situation, correspondence with cloud servers will present
additional postponements.[37–​40].

14.9 Research Scope
To facilitate the mix of AI/​ML, research endeavors are required in a few ways.
For example, the speed increase of a profound neural organization has to be
cutting edge for equal registering, quicker calculation, and distributed com-
puting. Appropriated profound learning frameworks present a chance for
5G to assemble the insight in its frameworks to convey high throughput and
super low idleness. There have been a few ongoing endeavors in increasing
speed of a profound neural organization.
5G-​IoT security and protection need more examination in the spaces of con-
firmation, approval, access control, and protection safeguarding. The current
3GPP-​defined networks utilize practical hub detail and conceptual interfaces
yet in 5G-​IoT. The actual organization will fill in as a center framework and
security affirmation will be the vital test to manage. At this stage, semi-​
managed AI-​helped arrangements better suit the appropriate frameworks.
With the advancement of AI calculations/​algorithms, these frameworks will
turn out to be completely mechanized later on.
252

252 Machine Learning Algorithms and Applications in Engineering

14.10 Conclusion
In this chapter, we focused on the integration of AI/​ML for 5G network
with respect to different constraints. We concentrated on a few contextual
investigations counting regulation characterization, channel coding, enor-
mous MIMO, reserving, energy proficiency, and network safety. As a con-
clusion of this top-​to-​bottom review, AI-​empowered 5G communication and
systems administration is a promising arrangement that can furnish remote
organizations with the insight, effectiveness, and adaptability needed to deal
with the alarm radio asset well and convey superior grade of administration
to the clients. In any case, a few endeavors are as yet expected to decrease
the intricacy of profound adapting so it tends to be executed in time-​delicate
organizations and low-​power gadgets and test the models in more reason-
able situations. These days, with the power and pervasiveness of data, various
specialists are adjusting their insight and growing their instruments arms
stockpile with AI-​based models, calculations, and practices, particularly in the
5G world, where even a couple of milliseconds of inactivity can have an effect.

References
[1] C. Zhang, P. Patras, and H. Haddadi, “Deep learning in mobile and wireless
networking: A survey,” IEEE Communications Surveys & Tutorials, 2019.
[2] M. Soltani, V. Pourahmadi, A. Mirzaei, and H. Sheikhzadeh, “Deep learning-​
based channel estimation,” IEEE Communications Letters, vol. 23, no. 4, pp. 652–​
655, 2019.
[3] S. Gao, P. Dong, Z. Pan, and G. Y. Li, “Deep learning based channel estimation
for massive MIMO with mixed-​resolution ADCS,” IEEE Communications Letters,
vol. 23, no. 11, pp. 1989–​1993, 2019.
[4] H. Ye, G. Y. Li, and B.-​H. Juang, “Power of deep learning for channel estimation
and signal detection in ofdm systems,” IEEE Wireless Communications Letters,
vol. 7, no. 1, pp. 114–​117, 2017.
[5] Devasis Pradhan and K.C. Priyanka, “RF-​energy harvesting (RF-​EH) for sustain-
able ultra dense green network (SUDGN) in 5G green communication,” Saudi
Journal of Engineering and Technology 2415–​6264 (Online, 5(6), pp. 2582013264
DOI: 10.36348/​sjet.2020.v05i06.001, 2020)
[6] Devasis Pradhan and K.C. Priyanka, “A comprehensive study of renewable
energy management for 5G green communications: Energy saving techniques
and its optimization,” Journal of Seybold Report 1533–​ 9211, 25(10), pp. 270–​
284. 2020.
[7] Devasis Pradhan and A. Dash, “An overview of beam forming techniques
toward the high data rate accessible for 5G networks,” International Journal
of Electrical, Electronics and Data Communication, 2320–​2084, 2321–​2950, 8(12),
pp. 1–​5, 2020.
253

Integration of AI/Ml in 5G Technology 253

[8] Devasis Pradhan and R. Rajeswari, “5G-​green wireless network for communica-
tion with efficient utilization of power and cognitiveness,” in Jennifer S. Raj (ed.),
International Conference on Mobile Computing and Sustainable Informatics.
Springer Nature Switzerland AG 2021:Springer, Cham, pp. 325–​335. 2020.
[9] Devasis Pradhan, P. K. Sahu, A. Dash, and Hla Myo Tun, “Sustainability of 5G
green network toward D2D communication with RF-​energy techniques,” 2021
International Conference on Intelligent Technologies (CONIT), 2021, pp. 1–​10,
doi:10.1109/​CONIT51480.2021.9498298
[10] Y. Wang, M. Narasimha, and R. W. Heath, “Mmwave beam prediction with situ-
ational awareness: A machine learning approach,” in 2018 IEEE 19th International
Workshop on Signal Processing Advances in Wireless Communications
(SPAWC). IEEE, 2018, pp. 1–​5.
[11] E. Balevi and J. G. Andrews, “Deep learning-​based channel estimation for high-​
dimensional signals,” arXiv preprint arXiv:1904.09346, 2019.
[12] H. He, C.-​K. Wen, S. Jin, and G. Y. Li, “Deep learning-​based channel estimation
for beamspace mmwave massive mimo systems,” IEEE Wireless Communications
Letters, vol. 7, no. 5, pp. 852–​855, 2018.
[13] M. S. Safari and V. Pourahmadi, “Deep ul2dl: Channel knowledge transfer from
uplink to downlink,” arXiv preprint arXiv:1812.07518, 2018.
[14] C.-​K. Wen, W.-​T. Shih, and S. Jin, “Deep learning for massive mimo csi feed-
back,” IEEE Wireless Communications Letters, vol. 7, no. 5, pp. 748–​751, 2018.
[15] M. Khani, M. Alizadeh, J. Hoydis, and P. Fleming, “Adaptive neural signal
detection for massive mimo,” arXiv preprint arXiv:1906.04610, 2019.
[16] G. Gao, C. Dong, and K. Niu, “Sparsely connected neural network for massive
MIMO detection,” EasyChair, Tech. Rep., 2018.
[17] M. Alrabeiah and A. Alkhateeb, “Deep learning for TDD and FDD
massive MIMO: Mapping channels in space and frequency,” arXiv preprint
arXiv:1905.03761, 2019.
[18] C. Zhang, P. Patras, and H. Haddadi, “Deep learning in mobile and wireless
networking: A survey,” IEEE Communications Surveys Tutorials, vol. 21, pp. 2224–​
2287, 2019.
[19] L. Huang, A. D. Joseph, B. Nelson, B. I. Rubinstein, and J. D. Tygar, “Adversarial
machine learning,” in Proceedings of the 4th ACM workshop on Security and
artificial intelligence, pp. 43–​58, ACM, 2011.
[20] Devasis Pradhan, P.K. Sahu, K. Rajeswari, and Hla Myo Tun, “A study of local-
ization in 5G green network (5G-​GN) for futuristic cellular communication,”
The 3rd International Conference on Communication, Devices and Computing
(ICCDC 2021), India, 16–​18 August 2021.
[21] G. A. Plan, “An inefficient truth.” Global Action Plan Report, 2007.
[22] Hla Myo Tun, “Radio network planning and optimization for 5G telecom-
munication system based on physical constraints”, Journal of Computer Science
Research, 3(1), January 2021. https://​doi.org/​10.30564/​jcsr.v3i1.2701
[23] W. Song, F. Zeng, J. Hu, Z. Wang, and X. Mao, “An unsupervised learning-​based
method for multi-​hop wireless broadcast relay selection in urban vehicular
networks,” IEEE Vehicular Technology, 865 Conference, vol. 2017.
[24] Pradhan, D., Sahu, P. K., Dash, A., & Tun, H. M. (2021, June). Sustainability of
5G Green Network toward D2D Communication with RF-​Energy Techniques.
In 2021 International Conference on Intelligent Technologies (CONIT) (pp. 1–​10). IEEE.
254

254 Machine Learning Algorithms and Applications in Engineering

[25] M. S. Parwez, D. B. Rawat, and M. Garuba, “Big data analytics for user-​activity
analysis and user-​ anomaly detection in mobile wireless network,” IEEE
Transactions on Industrial Informatics, vol. 13, no. 4, pp. 2058–​2065, 2017.
[26] L.-​C. Wang and S. H. Cheng, “Data-​driven resource management for ultra-​dense
small cells: An affinity propagation clustering approach,” IEEE Transactions on
Network Science and Engineering, vol. 4697, no. c, pp. 1–​1, 2018.
[27] U. Challita, L. Dong, and W. Saad, “Deep learning for proactive resource allo-
cation in LTE-​U networks,” in European Wireless 2017, 23rd European Wireless
Conference, 2017.
[28] R. Pascanu, T. Mikolov, and Y. Bengio, “On the difficulty of training recurrent
neural networks,” Tech. Rep., 2013.
[29] Q.V. Le, N. Jaitly, and G.E. Hinton Google, “A simple way to initialize recurrent
networks of rectified linear units,” Tech. Rep., 2015.
[30] G. Alnwaimi, S. Vahid, and K. Moessner, “Dynamic heterogeneous learning
games for opportunistic access in LTE-​based macro/​femtocell deployments,”
IEEE Transactions on Wireless Communications, vol. 14, no. 4, pp. 2294–​2308, 2015.
[31] D. D. Nguyen, H. X. Nguyen, and L. B. White, “Reinforcement learning with
network-​assisted feedback for heterogeneous RAT selection,” IEEE Transactions
on Wireless Communications, vol. 16, no. 9, pp. 6062–​6076, 2017.
[32] H. Sun, X. Chen, Q. Shi, M. Hong, X. Fu, and N. D. Sidiropoulos, “Learning to
optimize: Training deep neural networks for interference management,” IEEE
Transactions on Signal Processing, vol. 66, no. 20, pp. 5438–​5453, 2018.
[33] Y. Junhong and Y. J. Zhang, “Drag: Deep reinforcement learning based base
station activation in heterogeneous networks,” IEEE Transactions on Mobile
Computing, vol. 19, no. 9, pp. 2076–​2087, 2019.
[34] J. Liu, B. Krishnamachari, S. Zhou, and Z. Niu, “Deepnap: Data-​driven base
station sleeping operations through deep reinforcement learning,” IEEE Internet
of Things Journal, vol. 5, no. 6, pp. 4273–​4282, 2018.
[35] M. Kozlowski, R. McConville, R. Santos-​Rodriguez, and R. Piechocki, “Energy
efficiency in reinforcement learning for wireless sensor networks,” arXiv preprint
arXiv:1812.02538, 2018.
[36] B. Matthiesen, A. Zappone, E. A. Jorswieck, and M. Debbah, “Deep learning for
optimal energy-​efficient power control in wireless interference networks,” arXiv
preprint arXiv:1812.06920, 2018.
[37] C. Gutterman, E. Grinshpun, S. Sharma, and G. Zussman, “Ran resource
usage prediction for a 5G slice broker,” in Proceedings of the Twentieth ACM
International Symposium on Mobile Ad Hoc Networking and Computing.
ACM, 2019, pp. 231–​240.
[38] M. Chen, W. Saad, C. Yin, and M. Debbah, “Echo state networks for pro-
active caching in cloud-​based radio access networks with mobile users,” IEEE
Transactions on Wireless Communications, vol. 16, no. 6, pp. 3520–​3535, 2017.
[39] X. Lu, L. Xiao, C. Dai, and H. Dai, “Uav-​ aided cellular communications
with deep reinforcement learning against jamming,” arXiv preprint
arXiv:1805.06628, 2018.
[40] M. Sadeghi and E. G. Larsson, “Physical adversarial attacks against end-​to-​end
autoencoder communication systems,” IEEE Communications Letters, vol. 23, no.
5, pp. 847–​850, 2019.
newgenprepdf

255

15
Electrical Price Prediction using
Machine Learning Algorithms

Swastik Mishra, Kanika Prasad,* and Anand Mukut Tigga


Department of Production and Industrial Engineering, National Institute of Technology,
Jamshedpur, India
*[email protected]

CONTENTS
15.1 Introduction.............................................................................................. 256
15.2 Literature Review..................................................................................... 257
15.3 Methodology............................................................................................. 258
15.3.1 Univariate Forecasting Models................................................. 258
15.3.1.1 Autoregressive Model............................................... 258
15.3.1.2 Moving Average (MA).............................................. 259
15.3.1.3 ARMA.......................................................................... 259
15.3.1.4 ARIMA........................................................................ 259
15.3.1.5 LSTM........................................................................... 260
15.3.2 Multivariate Forecasting Models............................................. 260
15.3.2 1 DNN............................................................................ 260
15.3.2.2 CNN............................................................................. 261
15.3.2.3 CNN-LSTM................................................................. 261
15.3.2.4 CNN-LSTM-DNN..................................................... 261
15.4 Electrical Demand Forecasting............................................................... 261
15.4.1 Dataset.......................................................................................... 261
15.4.2 Data Preprocessing..................................................................... 262
15.4.3 Experiment.................................................................................. 262
15.4.3.1 Univariate Time-Series Forecasting........................ 262
15.4.3.2 Multivariate Time-Series Forecasting..................... 264
15.5 Results and Discussions.......................................................................... 264
15.5.1 For Univariate Models............................................................... 264
15.5.2 For Multivariate Models............................................................ 266
15.6 Conclusions............................................................................................... 267
15.7 Limitations and Future Scope................................................................ 268

DOI: 10.1201/9781003104858-15 255


256

256 Machine Learning Algorithms and Applications in Engineering

15.1 Introduction
Energy is essential for the economic development of any country. In the
case of developing countries, the energy sector assumes critical importance
given the ever-​increasing energy needs requiring huge investments to meet
them. Electricity has influenced our day-​to-​day life and in industries, electri-
city provides power to large machines, which produce essential utilities and
industrial and consumer products. Electricity provides the basis for the eco-
nomic development of any country because a consistent and reliable power
source creates a variety of businesses and job opportunities. It provides access
to online resources and information.
Electricity is a secondary form of energy, because it’s generated from the
conversion of energy produced from different sources such as renewable,
for example, solar energy, hydro energy, onshore offshore energy, and non-
renewable such as coal, lignite, natural gases, and fossil fuels. It cannot be
stored in large quantities due to its properties. Sometimes, there is a disbalance
in the generation and supply of electrical energy. This causes fluctuation in
the electricity prices. Extreme price volatility has forced power generators to
hedge not only volume but also price risk. Price forecasts from a few hours to
a few months ahead have been of particular interest to power portfolio man-
agers. A power generation company capable of predicting volatile wholesale
prices with a reasonable level of accuracy can adjust its strategy and its own
production or consumption schedule to reduce the risk or maximize profits
in day-​ahead trading. It helps management take the right decisions and
manage resources accordingly and work on the shortage. Lack of sufficient
power generation capacity, poor transmission and distribution infrastruc-
ture, high costs of supply to remote areas, or simply lack of affordability for
electricity are among the biggest hurdles for extending grid-​based electricity.
Forecasting of power price plays an essential role in the electricity industry,
as it provides the basis for making decisions in power system planning and
operation.
Electricity price forecasting is a time-​ series forecasting problem.
Therefore, in this research work, electricity prices are forecast by applying
machine learning (ML) algorithms with an open-​source data set. The clas-
sical forecasting models and the combination of recurring neural networks
(RNN) and convolutional neural networks (CNN) are used for price pre-
diction. Classical models such as moving average, autoregressive moving
average (ARMA), autoregressive integrated moving average (ARIMA),
and long short-​term memory (LSTM) are used for electricity price predic-
tion. Electricity prices are governed by different variables known as price
determinants. Therefore, after the application of classical forecasting models
and RNN, multivariate forecasting models are applied on the same dataset.
Consequently, the results obtained from the models applied are compared
257

Electrical Price Prediction using Machine Learning Algorithms 257

based on mean absolute percentage error (MAPE) and mean absolute error
(MAE) values for both univariate models and multivariate models. The
graphs of the actual price and forecasting price are also plotted to visualize
the results obtained.

15.2 Literature Review
Researchers have already applied ML algorithms for forecasting the change
in prices. Rolnick et al. defined the climate change problem and encouraged
the use of ML algorithms for forecasting the adverse effects of climate change
[1]‌. Kof Nti et al. presented a review on the classification of electrical energy
and categorized the forecasting methods into industrial and artificial intelli-
gence (AI) methods [2]. Aishwarya et al. applied ML models such as Bayesian
linear regression, boosted decision tree regression, decision forest regression,
and statistical analysis method for predicting the number of customers with
internal data and external data in a ubiquitous environment [3]. Mir et al.’s
comparative review shows that the time-​series modeling approach has been
extensively used for long-​and medium-​term forecasting. AI-​based techniques
for short-​term forecasts remain prevalent in the literature [4]. Johannesen
et al. compared various regression tools based on the lowest MAPE value
and concluded that random forest regressor provides better short-​term load
prediction (30 min) and K-​nearest neighbor (KNN) offers relatively better
long-​term load prediction (24 h) [5]. Ahmad and Chen determined that wea-
ther change is responsible for the change in energy consumption pattern in
the domestic, commercial, and industrial sectors and concluded that LSBoost
performance is more modest than the LMSR and NARM for monthly, sea-
sonally, and yearly ahead intervals [6]. Fattah et al. developed an ARIMA
model for demand forecasting of a product using the Box–​Jenkins time-​series
approach [7]. For understanding the nature of time series and the objective
of analysis, Athiyarath et al. explained and applied the concepts of LSTM,
ARIMA, and CNN [8]. Kaushik et al. evaluated different statistical, neural,
and ensemble techniques in their ability to predict patients’ weekly average
expenditures on certain pain medications. Two statistical models, persistence
(baseline) and ARIMA, multilayer perceptron model, LSTM model and an
ensemble model were used to predict the expenditures on two different pain
medications [9]. Ahmad et al. used artificial neural networks (ANN) with
a nonlinear autoregressive exogenous multivariable inputs model, multi-
variate linear regression model, and adaptive boosting model to predict
energy demand in a smart grid environment [10]. Allee et al. used a data-​
driven approach for demand prediction using survey and smart meter data
from 1,378 Tanzanian mini-​grid customers. Applied support vector machines
258

258 Machine Learning Algorithms and Applications in Engineering

were used to predict building energy consumption in tropical regions [11].


Dong et al. presented support vector machine (SVM), a new neural net-
work algorithm, to forecast energy consumption in buildings [12]. Jamii, an
ARIMA model, was applied to model electrical energy consumption annu-
ally from 1971 to 2020 [13]. Hasanah et al. compared the performance of two
methods in electric load demand forecasting. Genetic algorithm–​support
vector machine (GA–​SVM) and ARIMA methods were applied for the pre-
diction of daily load in Malang city, Indonesia [14]. Aurna et al. used ARIMA
and Holt Winters model for the energy consumption of Ohio and Kentucky
states and predicted the consumption considering different periods (daily,
weekly, monthly) [15]. Velasco et al. used ARIMA and ANN as forecasting
models to predict day-​ahead electric load. Electric load data preparation,
model implementation, and forecasting evaluation were done to assess if the
prediction of the models met the acceptable error tolerance [16].
It is evident from the literature review that most of the research works
have applied univariate forecasting models for price prediction. Since elec-
trical price prediction is a multivariate forecasting problem, in this work,
different univariate and multivariate forecasting models are applied. The
results obtained are compared using MAE and MAPE metrics to determine
the accuracy of prediction.

15.3 Methodology
This section presents the various univariate and multivariate forecasting
models that were applied in this work.

15.3.1 Univariate Forecasting Models


15.3.1.1 Autoregressive Model
An autoregressive (AR) model predicts the future behavior based on past
behavior. AR models are also called conditional models, Markov models,
or transition models. They are used for forecasting when there is some cor-
relation between values in a time series and the values that precede and
succeed them.
An AR model is represented by Equation 15.1, where the value of outcome
variable (P) at some point t in time is like “regular” linear regression, dir-
ectly related to the predictor variable (Q). The simple linear regression and
AR models differ in this respect that P is dependent on Q and also previous
values of P.

Pt =​ δ +​ φ1Pt–​1 +​ φ2Pt–​2 +​… +​φpPt–​p +​ At (15.1)


259

Electrical Price Prediction using Machine Learning Algorithms 259

where
• Pt–​1, Pt–​2…Pt–​p are the past series values (lags),
• At is white noise (i.e., randomness), and
• δ is defined by the following equation:

δ = ​(1 – ​∑ φi)µ (1≤ i ≤p) (15.2)

15.3.1.2 Moving Average (MA)


MA is one of the most foundational models for time-​ series forecasting,
denoted as MA(a) where a is the order of the model. This is one of the basic
statistical models that is a building block of more complex models. The MA
model states that the current value is linearly dependent on the current and
past error terms. Again, the error terms are assumed to be mutually inde-
pendent and normally distributed, just like white noise. Equation 15.3
represents the predicted value using MA technique:

Pt =​ µ+​ α1 +​ φ1 α t–​1 +​ φ2 α t–​2+​ … +​φa α t–​a (15.3)

where Pt =​present value


µ =​mean of the series
φ =​The magnitude of the impact of past errors on the present value
α =​the present error term, and past error terms.
a =​order of the series

15.3.1.3 ARMA
An ARMA model is used to describe weakly stationary stochastic time
series in terms of two polynomials. The first of these polynomials is for
autoregression, and the second for the MA. In this model, the impact of past
lags along with the residuals is considered for predicting the future values as
explained in the following equation:

Pt =​ β₁* P​t–​1 +​ α₁* Ɛt–​1 +​ β₂* Pt–​2 +​ α₂*Ɛt–​2 +​ β₃*Pt–​3


+​ α₃*Ɛt–3 +​ … +​ β*Pt–​k​+​ α*Ɛt–​k​ (15.4)

where
β =​the coefficients of the AR model
α =​the coefficients of the MA model.

15.3.1.4 ARIMA
ARIMA is a class of models that explain a given time series data based on its
own previous values, that is, its own lags and the lagged forecast errors, so
260

260 Machine Learning Algorithms and Applications in Engineering

that Equations 15.5–​15.7 can be used to forecast future values. The ARIMA
model is used when the mean of the time series is not constant, that is, the
time series is not stationary. Therefore, the data is prepared by a degree of dif-
ferencing in order to make it stationary, that is, to remove trend and seasonal
structures that negatively affect the regression model:

Zt=​ at+​1 –​ at (15.5)

Zt =​ ϕ₁*Zt–​1 +​ ɵ1*€t–​1 +​€t (15.6)

a1 = ∑ i = 1 Zk − i + a1
k =1
(15.7)

where
at ​ = data points on the non-​stationary time series
ϕ₁*Zt–​1 =​represents the autoregressive part
ɵ1*€t–​1 =​represents the moving average part
€t =​error occurred

15.3.1.5 LSTM
LSTMs are special kind of RNNs, precisely designed to avoid the long-​term
dependency in problems. They can remember information for longer period
of time with ease and it is their intended property. They possess a sigmoid
neural net layer and a pointwise multiplication operation. The sigmoid layer
outputs numbers between 0 and 1, describing how much of each component
should be let through. A value of 0 means “let nothing through,” while a value
of 1 means “let everything through!” The vanishing gradient effect is red by
implementing three gates along with the hidden state. The three implemented
gates are commonly referred to as input, output, and forget gates. An LSTM
has these gates to protect and control the cell state. Efficient applications of
LSTM networks can be found in research fields such as human trajectory pre-
diction, traffic forecasting, speech recognition, and weather prediction.

15.3.2 Multivariate Forecasting Models


15.3.2.1 DNN
DNNs are enhanced versions of the conventional neural networks. They are
becoming very popular in recent times due to their outstanding performance
to learn not only the nonlinear input–​output mapping but also the under-
lying structure of the input data vectors. These types of models can approxi-
mate the behavior of any function (universal approximation theorem). The
DNN output can be explained by the following equation:


pil = f (∑ J
j =1
(w q + bi
i,k k )) (15.8)
261

Electrical Price Prediction using Machine Learning Algorithms 261

where p is the output of neuron of the current layer, q is the output of neuron
of the previous layer, w is the set of weights, b is bias, f is nonlinear activa-
tion, i is current neuron, l is current layer, and k is the earlier layer.

15.3.2.2 CNN
CNN is traditionally designed for image detection systems but can also be
used as time-​series forecasting model. The neurons in CNN can hold high
dimensional data and so this model performs better than the other neural
networks. CNN models have three prominent layers: (i) convolutional layer,
(ii) pooling layer, and (iii) flatten layer. A flatten layer is used between the
convolutional layers and the dense layer to reduce the feature maps to a
single one-​dimensional vector.

15.3.2.3 CNN-​LSTM
This architecture is also known as long-​term recurrent convolutional net-
work (LRCN) model. The CNN-​LSTM architecture uses CNN layers for fea-
ture extraction on input data combined with LSTMs to support sequence
prediction. A CNN-​LSTM can be defined by adding CNN layers at the front
end followed by LSTM layers with a dense layer at the output. These models
were developed for visual time-​series prediction problems.

15.3.2.4 CNN-​LSTM-​DNN
In order to perform complex tasks, the neural networks need to be trained
deeper. When deeper networks start to converge, the problem of degrad-
ation occurs. In degradation, with the increase in network depth, accuracy
gets saturated and then decreases rapidly. Unexpectedly, such degradation is
not caused by overfitting, and adding more layers to a suitably deep model
leads to higher training error. Skip connections, which are also known as
shortcut connections, are introduced to overcome this problem. This type
of connection skips some of the layers in the neural network and feeds the
output of one layer as the input to the next layers. In this CNN-​LSTM model,
output is skipped to common DNN.

15.4 Electrical Demand Forecasting


15.4.1 Dataset
The dataset used in the study contains four years of electrical consumption,
generation, pricing, and weather data for Spain. Consumption and gener-
ation data were retrieved from ENTSOE, a public portal for transmission
262

262 Machine Learning Algorithms and Applications in Engineering

service operator (TSO) data. The dataset is unique because it contains hourly
data for electrical prices and the respective forecasts by the TSO for consump-
tion and pricing. The work focuses on predicting electric prices better than
that already forecast by the present data. The metrics we are using for com-
parison is MAPE and MAE defined below:

100% n At − Ft
MAPE =
n
∑ t =1 At (15.9)


n
i =1
At − ft
MAE = (15.10)
n

where
At =​actual value
Ft =​forecast value
n =​total number of datapoints

15.4.2 Data Preprocessing
In order to apply ML algorithms, the raw dataset is first cleaned and
preprocessed. After cleaning the data, that is eliminating all the null values
and noises, the dataset is converted to a usable format. The next step is fea-
ture scaling, after which the data is split into testing and training dataset.
For univariate time-​series forecasting, a single feature is extracted and the
value for total actual price is predicted. Subsequently, a min–​max scaler is
employed to scale the dataset. The dataset is divided into two parts: 80% as a
training dataset and 20% as a testing dataset.
For multivariate time-​series forecasting, multiple features, that is, energy
consumption, price, day of the week, and month of the year are extracted.
Then the scaling of the features is done using min–​max scaler.

15.4.3 Experiment
15.4.3.1 Univariate Time-​Series Forecasting
After data preprocessing and feature selection, the selected forecasting
models are successfully applied using Jupyter notebook. Jupyter notebook
is an open-​source workbench used to create and share documents having
codes, projects, equations, and visualization. The forecasted and actual
prices are plotted in the graphs shown in Figure 15.1. From the figure, it
can be concluded that the forecasted price is closest to actual price for the
LSTM model. MAPE and MAE values obtained for each model is shown in
Table 15.1.
263

Electrical Price Prediction using Machine Learning Algorithms 263

FIGURE 15.1
The actual and the forecasted price is plotted in the graph, predicted by the univariate
forecasting models.
264

264 Machine Learning Algorithms and Applications in Engineering

TABLE 15.1
The MAPE, RMSE, and MAE Values Obtained
for Each Univariate Model

Models MAPE MAE

Autoregressive 20.0674 12.4224


MA 11.7187 0.0484
ARMA 12.4534 20.4573
ARIMA 84.8715 55.6936
LSTM 4.1564 2.6708

15.4.3.2 Multivariate Time-​Series Forecasting


In this section, all the multivariate forecasting models that are applied are
explained. All the models have used the same final two DNN layers with
dropout.
All the unique layers of all the models are stated below:

1. A three-​layer DNN (one layer plus the common bottom two layers)
2. A CNN with two layers of 1D convolutions with max pooling
3. A LSTM with two LSTM layers
4. A CNN-​stacked LSTM with layers from models 2 and 3 feeding into the
common DNN layer. We are using the same layers from the CNN and
LSTM model, stacking the CNN as input to the pair of LSTMs.
5. A CNN is stacked LSTM with a skip connection to the common
DNN layer.

The same CNN and LSTM layers as the previous models are used this time
with a skip connection direct to the common DNN layer.
Figure 15.2 shows that the LSTM appears to oscillate over a longer fre-
quency compared to the other models. The CNN also seems to capture the
intraday oscillations. CNN-​stacked LSTM shows how these two attributes
of the model’s learning are combined. Table 15.2 shows the MAE values
obtained by the different models.

15.5 Results and Discussions


15.5.1 For Univariate Models
After analyzing the results obtained after the application of the forecasting
models, the LSTM forecasting model gives the most accurate results in
newgenrtpdf
265
Electrical Price Prediction using Machine Learning Algorithms
FIGURE 15.2
The prediction of price for two weeks using multivariate forecasting models.

265
266

266 Machine Learning Algorithms and Applications in Engineering

TABLE 15.2
The MAPE Values Obtained by the Different Multivariate Models

MODELS DNN CNN CNN-​LSTM stacked CNN-​LSTM skip

MAPE 25.7879 21.4385 22.7875 23.6760


MAE 0.0877 0.0789 0.0828 0.0882

FIGURE 15.3
Comparison between the applied univariate models on the basis of MAE.

comparison to all the other forecasting models. If MAPE is taken as the metric
of comparison, LSTM model gives the lowest MAPE. Lower MAPE shows
that the LSTM forecasting results are closer to the actual prices as compared
to other models. Figure 15.4 shows the MAPE values of all univariate models
and it can be seen that MAPE of ARIMA is maximum, that is, ARIMA model’s
accuracy is the lowest. Figure 15.1 shows the comparison of actual prices
and forecasted prices, and it can be seen that the prices forecasted by LSTM
overlap with the actual price graph. Figure 15.3 shows the MAE obtained
from all the univariate models applied and it can be seen MAE value for
ARIMA is maximum.

15.5.2 For Multivariate Models


The CNN-​stacked LSTM model gives the best results in comparison to LSTM
skipped method. CNN and DNN also performed exceptionally well. The
models are compared on the basis of MAPE and MAE metrics. Figure 15.5
shows the MAPE values of all models and it can be seen MAPE of CNN-​
stacked LSTM is lower as compared to LSTM skipped and MAPE of CNN is
lowest. Figure 15.6 shows the MAE values obtained from the applied multi-
variate models. The MAE value obtained for CNN method is lowest among
all the methods and therefore this model performed really well. The MAE
267

Electrical Price Prediction using Machine Learning Algorithms 267

FIGURE 15.4
Comparison between the applied univariate models on the basis of MAPE.

FIGURE 15.5
Comparison between the applied multivavariate models on the basis of MAPE.

of LSTM skipped is more than that of CNN stacked and that’s why the fore-
casted value obtained from CNN stacked is close to actual one in comparison
to CNN skipped.

15.6 Conclusions
Electricity plays a key role in the economic development of any nation. All
the power generation companies in the market aim to utilize all the available
268

268 Machine Learning Algorithms and Applications in Engineering

FIGURE 15.6
Comparison between the applied multivavariate models on the basis of MAE.

resources to the fullest, and improve the generation, consumption, and supply
of electricity produced in order to maximize profit generation. Electricity price
prediction plays a vital role in accomplishing the aims of the power companies.
The accurate prediction of the electric prices enables the power managers to
make decisions regarding the raw materials used and the production and
supply of electricity. In this chapter, the idea is to apply different univariate
and multivariate machine learning algorithms for electricity price prediction.
Jupyter notebook, an open-​source workbench, is used for coding and visual-
izing the results. Electric price predictions involve historic data of factors such
as load and time in hours, days, and weeks. The machine learning algorithms
perform well and handle a large amount of data. Out of the univariate models,
LSTM outperforms all the other methods. Out of multivariate models, LSTM
skip performs better than LSTM stacked. CNN and DNN give satisfactory
results. As managers aim to understand the dynamics of the market, the
methods discussed in this chapter can be used by them to create a concrete
basis to make their decisions. Lower errors in forecasting lead to accurate
decisions to be made in resource planning and capacity planning, which in
turns improves the profit share of the power generation companies.

15.7 Limitations and Future Scope


In this research, the impact of weather on the forecasting model is not
incorporated. Also, same techniques may be applied for electrical energy
price forecasting in Indian context once the dataset is obtained.
269

Electrical Price Prediction using Machine Learning Algorithms 269

References
[1] Rolnick, D., Donti, P.L., Kaack, L.H., Kochanski, K., Lacoste, A., Sankaran,
K., Ross, A.S., Milojevic-​ Dupont, N., Jaques, N., Waldman-​ Brown, A., and
Luccioni, A., 2019. Tackling climate change with machine learning. arXiv pre-
print arXiv:1906.05433.
[2] Nti, I.K., Teimeh, M., Nyarko-​Boateng, O., and Adekoya, A.F., 2020. Electricity
load forecasting: a systematic review. Journal of Electrical Systems and Information
Technology, 7(1), 1–​19.
[3] Aishwarya, K., Aishwarya Rao, Nikita Kumari, Akshit Mishra, and Rashmi,
M.R., 2020. Food demand prediction using machine learning, International
Research Journal of Engineering and Technology, 7(6), 3672–​3675.
[4] Mir, A.A., Alghassab, M., Ullah, K., Khan, Z.A., Lu, Y., and Imran, M., 2020.
A review of electricity demand forecasting in low and middle income coun-
tries: The demand determinants and horizons. Sustainability, 12(15), 5931.
[5] Johannesen, N.J., Kolhe, M., and Goodwin, M., 2019. Relative evaluation of
regression tools for urban area electrical energy demand forecasting. Journal of
Cleaner Production, 218, 555–​564.
[6] Ahmad, T. and Chen, H., 2019. Nonlinear autoregressive and random forest
approaches to forecasting electricity load for utility energy management
systems. Sustainable Cities and Society, 45, 460–​473.
[7] Fattah, J., Ezzine, L., Aman, Z., El Moussami, H., and Lachhab, A., 2018.
Forecasting of demand using ARIMA model. International Journal of Engineering
Business Management, 10, 1–​9.
[8] Athiyarath, S., Paul, M., and Krishnaswamy, S., 2020. A comparative study and
analysis of time series forecasting techniques. SN Computer Science, 1, 1–​7.
[9] Kaushik, S., Choudhury, A., Sheron, P.K., Dasgupta, N., Natarajan, S., Pickett,
L.A., and Dutt, V., 2020. AI in healthcare: time-​series forecasting using statistical,
neural, and ensemble architectures. Frontiers in Big Data, 3, 4.
[10] Ahmad, T. and Chen, H., 2018. Potential of three variant machine-​learning
models for forecasting district level medium-​ term and long-​ term energy
demand in smart grid environment. Energy, 160, 1008–​1020.
[11] Allee, A., Williams, N.J., Davis, A., and Jaramillo, P., 2021. Predicting initial elec-
tricity demand in off-​grid Tanzanian communities using customer survey data
and machine learning models. Energy for Sustainable Development, 62, 56–​66.
[12] Dong, B., Cao, C., and Lee, S.E., 2005. Applying support vector machines to pre-
dict building energy consumption in tropical region. Energy and Buildings, 37(5),
545–​553.
[13] Mohammed, J. and Mohamed, M., 2021. The forecasting of electrical energy
consumption in morocco with an autoregressive integrated moving average
approach, Hindawi Mathematical Problems in Engineering, 2021, Article ID
6623570, 9.
[14] Hasanah, R.N., Indratama, D., Suyono, H., Shidiq, M., and Abdel-​Akher, M.,
2020. Performance of genetic algorithm-​support vector machine (GA-​SVM) and
autoregressive integrated moving average (ARIMA) in electric load forecasting.
Journal FORTEI-​JEERI, 1(1), 60–​69.
[15] Aurna, Md. N.F., Rubel, T.M., Siddiqui, T.A., Karim, T., Saika, S., Md. Arifeen,
M., Mahbub, T.N., Salim Reza, S.M., and Kabir H. 2021. Time series analysis of
270

270 Machine Learning Algorithms and Applications in Engineering

electric energy consumption using autoregressive integrated moving average


model and Holt Winters model, TELKOMNIKA Telecommunication, Computing,
Electronics and Control 19(3), 991–​1000.
[16] Velasco, L.C.P., Polestico, D.L.L., Macasieb, G.P.O., Reyes, M.B.V., and Vasquez
Jr, F.B., 2018. Load forecasting using autoregressive integrated moving average
and artificial neural network. International Journal of Advanced Computer Science
and Applications, 9, 23–​29.
newgenprepdf

271

16
Machine Learning Application to Predict
the Degradation Rate of Biomedical Implants

Pradeep Bedi,1 Shyam Bihari Goyal,2 Prasenjit Chatterjee,3 and


Jugnesh Kumar4
1
Galgotais University, Greater Noida, Uttar Pradesh, India
2
City University, Petaling Jaya, Malaysia
3
MCKV Institute of Engineering, Liluah, Howrah West Bengal, India
4
St. Andrews Institute of Technology & Management, Gurgaon, Haryana, India
E-​mail: [email protected]; [email protected]; [email protected];
[email protected]

CONTENTS
16.1 Introduction................................................................................................. 271
16.2 Related Work............................................................................................... 275
16.3 Proposed Methodology............................................................................. 277
16.4 Conclusion................................................................................................... 280

16.1 Introduction
A bone fracture often termed as a broken bone, is a medical condition
in which the shape or contour of bone changes due to impact to external
forces or injuries under many biological as well as mechanical circumstances
such as injuries during physical activities, vehicle accidents, accidental falls
or due to weakening of bones because of aging as well an underlying disease
[1]‌. Under fracture conditions, broken or cracked bone is stabilized and
supported to handle the weight of the body for movement during the process
of fracture healing. Some fractures are healed from outside of the body
using plasters, but using them gives rise to critical issues. To reduce the risk
of infection from external support, surgical procedures are performed to
implant supports internally to stabilize fractured bones with some implants

DOI: 10.1201/9781003104858-16 271


272

272 Machine Learning Algorithms and Applications in Engineering

TABLE 16.1
Different Types of Fractures and Their Internal Fixators

Types of Fracture Internal Fixators

Skull fracture Wires, pins, and plates


Craniofacial fracture Wires, screws, and plates
Pelvic fracture Screws, plates, and external fixators
Spinal fracture Rods, pedicle screws, and plates
Upper and lowerlimb fracture Plates and nails

FIGURE 16.1
Bone healing procedure.

such as plates, screws, nails, or wires [2]. Different types of implant devices
used for fracture healing are shown in Table 16.1.
These implants are of two types: non-​biodegradable and biodegradable.
Non-​biodegradable implants have to be removed by second surgery after
healing from fracture. But biodegradable implants degrade the inside body
and therefore a second surgery is not needed after healing of the fractured
bones. Degradation of biological implant materials occurs by an electro-
chemical reaction in presence of an electrolyte, which results in the forma-
tion of oxides, hydroxides, hydrogen gas, or other compounds, as shown in
Figure 16.1. Metals are used as bone implant materials in a non-​biodegradable
implantation to fix cracked, deformed, worn-​out, or broken bones.
The artificial replacements are made up of metals, polymers, or ceramic
to be strong enough as well as flexible for everyday movement, as given
in Figure 16.2. So, while choosing a material, it is necessary to ascertain its
mechanical strength, flexibility, and biocompatibility. Biomaterials are syn-
thetic materials, either degradable or non-​degradable in nature, but they
must possess good load-​bearing capacity if they are to be used inside the
human body [3–​5].
Orthopedic biomaterials are implanted near the bone fracture to provide
support and to heal the bone tissues. The end of fractured bones is connected
to implant devices, which are fixed with metal pins or screws. After the
healing of fractured bones, these screws and pins are removed. If these screws
are made up of non-​biodegradable materials, after healing, they need to be
removed by a second operation.
273

Machine Learning Application to Predict the Degradation Rate 273

FIGURE 16.2
Bone fracture implant materials, issues, and research scope areas.

So, to avoid the second surgery, researchers are showing interest in


developing biodegradable materials that can be used as implants. The bio-
degradable screws, pins, nails, plates, and wires used as implants are gener-
ally left in the body after healing of bone fracture and they gradually
degrade inside the body [6]‌. As the materials are left inside the body, the
cytotoxic effect of biomedical materials have to be analyzed before they can
be employed as implant devices. Researchers are exploring biodegradable
materials that can improve the tissue regeneration process.
But firstly, it is necessary to understand what properties are essential for
artificial replacement devices or implants. The required properties are as
follows:

(i) The implants need to have a high biocompatibility.


(ii) The implants need to possess good mechanical properties, such as
strength, during the process of bone healing.
(iii) The implants should show good thermal conductivity.
(iv) The implants should cause least friction at the implant location.
(v) The implants need to be resistant to mechanical shock.

Implants have to necessarily possess the above-​mentioned properties during


the replacement process [7]‌. While designing any bone fracture implant
274

274 Machine Learning Algorithms and Applications in Engineering

device, its surface and mechanical properties have to be studied. If the selec-
tion of the material or its alloy composition is not accurate, then it may lead to
failure of the implant through loosening, osteolysis, wear, or toxic effect [8].
Subsequently, it is required to analyze the surrounding environment of
implant device and tissues where the device is placed. Therefore, for better
functioning of the biomedical implant, the choice of the appropriate material
holds high importance.
The selection of materials for the implantation is a crucial step for successful
long-​term implants. The study of the implant material and its properties need
to be focused upon for a successful treatment.
The characteristics of the implant materials studied are modulus of elasti-
city, compressive strength, tensile strength, shear strength, yield, and fatigue
strength, ductility, hardness, corrosive properties (crevice corrosion, pitting
corrosion, galvanic corrosion, electrochemical corrosion), surface tension,
surface energy, and surface roughness [8]‌. Table 16.2 shows the material
properties that can be considered for implantations.
The biodegradation progression is driven by three major factors such as
chemical, mechanical, and biological interactions. In the case of the chemical-​
based deterioration of the polymer, the degradation rate is highly dependent
on the polymer’s composition, crystallinity, molecular structure as well as its
hydrophobic and hydrophilic nature [2]‌.
The chemical degradation process is achieved by breaking the polymer’s
molecular chains, breaking the cross-​linking structure, or making interference

TABLE 16.2
Classification of Biomaterials

Material Types

Material Nature Metals Ceramics Polymers

Bio-​toxic Gold
Co-​Cr alloys Polyethylene
Stainless steel Polyamide
Niobium Polymethylmethacrylate
Tantalum Polytetrafluroethylene
Polyurethane
Bio-​inert Commercially pure Al oxide
titanium
Titanium alloy Zirconium oxide
(Ti-​6AL-​4U)
Bio-​active Hydroxyapatite
tricalcium phosphate
Bio-​glass
carbon silicon
Biodegradable Magnesium, Calcium phosphate, Silk, collagen, polylactic acid
Zinc, calcium, iron silica, alumina
275

Machine Learning Application to Predict the Degradation Rate 275

in its crystallinity. Bulk or surface degradation occurs in the body. In the case
of bulk degradation, the rate of degradation is attained at a faster rate as
in hydrophilic polymers by achieving their conversion into water-​soluble
materials. Surface degradation takes place in hydrophobic polymers in which
it is intended to keep the inner structure intact and offers better control over
degradation rates [3]‌.
In the case of degradation at the biological level, the materials are exposed
to the body fluids that result in changes in the chemical composition of the
polymers. The degradation can through enzymatic, oxidation, or hydrolytic
methods.
However, the interaction levels of the tissue and its behavior at the implant
site depend on its physical, biological, and chemical nature. On the basis of
their nature, implant device materials are categorized as:

Bio-​toxic: Due to this nature, the surrounding tissues of implanted


devices die.
Bioinert: Such materials show minimum interaction with surrounding
cells or tissues. They don’t show any adverse biological response.
Bioactive: Such materials are nontoxic in nature but show biological
responses by releasing some chemical ions.
Bioresorbable/​biodegradable: Such materials are nontoxic in nature and
get dissolved inside the body. These materials replace or regenerate
surrounding tissues biologically [5,6].

16.2 Related Work
Mahdi Dehestani et al. [9]‌experimentally investigated the mechanical prop-
erties and corrosion behavior of iron and hydroxyapatite (HA) composites
for biodegradable implant applications. It was observed that the mechan-
ical strength decreases with increasing HA content and decreasing HA par-
ticle size whereas their corrosion rates increased. Fe–​2.5 wt% HA was finally
created as the strongest composite.
Rakesh Rajan et al. [10] investigated the zinc–​ magnesium composite
implant material for its mechanical, corrosion, and biological properties
of magnesium. The mechanical strength is nearly that of the bone and the
corrosion rate was observed to be 0.38 mm/​year with a 12% elongation rate.
Richard et al. [11] investigated experimentally the corrosion properties
of Fe–​Mn–​Si alloys for biodegradable medical implants. A corrosion rate of
0.24–​0.44 mm/​year was observed.
Tong et al. [12] studied the microstructure, mechanical properties, bio-
compatibility, and degradation behavior of Zn–​Ge alloy for biodegradable
276

276 Machine Learning Algorithms and Applications in Engineering

implants. The experimental results show that it has a tensile strength of


about 53.9 MPa and corrosion rate of 0.1272 mm/​year, and elongation rate
of about 1.2%.
Watrob et al. [13] experimentally designed a Zn-​Ag-​Zr alloy with enhanced
strength as a potential biodegradable implant material that showed 17.1 ± 1.0
μm/​year corrosion rate and also inhibits bacterial growth.
Dandan Xia et al. [14] investigated an alloy of Mg-​Li-​Ca for bone implant
application, which was experimentally implemented on femurs of mice. This
alloy shows excellent biocompatibility.
Suryavanshi et al. [15] developed a bone screw using a metal–​polymer bio-
degradable composite for orthopedic applications. This was experimentally
analyzed on rats and no toxic effects were observed.
Zhang et al. [16] analyzed the mechanical properties of magnesium
nanocomposites, which were homogeneous exhibiting higher strength
and ductility. These two properties reduced the elongation percentage of
material.
Hongtao Yang et al. [17] prepared biodegradable zinc as biodegradable
bone implants for load-​bearing applications.
Razzaghi et al. [18] investigated the microstructure, mechanical properties,
and biocompatibility of nanoalloy for implant applications. A major increase
in the corrosion rate was observed.
Going by the literature, experimental analysis was undertaken for deter-
mining the suitable materials for implant designs and their properties. Some
researchers are focused on the application of a machine learning approach
for the optimization of performance parameters. Some contributions
in the field of application of machine learning and implant design are
discussed below.
Borgiani et al. [19] investigated and optimized the stress shielding effect of
the hip prosthesis to achieve a better performance of implants. The optimiza-
tion of the geometrical design of implants is achieved by machine learning
techniques such as artificial neural network (ANN) and support vector
machine (SVM). Parameters such as total stem length, thickness, and distance
between the implant neck and stem surface are used for optimization. These
parameters directly impact the stress shielding effect.
Chatterjee et al. [20] analyzed the strain deviation at different locations
inside the body under two conditions, that is, before and after implantation.
For the optimization of strain deviation, genetic algorithm was applied by
finding the co-​relation between the implant and their finite element simula-
tion data by using artificial neural network (ANN). Different bone conditions
were used to decide the specific guidelines while designing implant devices
that must be patient-​specific.
Niculescu et al. [21] studied the biomechanical behavior of ortho-
pedic implants by applying deep learning and support vector machine.
Further, mechanical testing was also performed for the analysis of
implant stiffness.
277

Machine Learning Application to Predict the Degradation Rate 277

Borjali et al. [22] and Bedi et al. [23] used machine learning methods to pre-
dict the wear rate of biomedical implants and to validate them by quantifying
the prediction error.

16.3 Proposed Methodology
The following problems are faced:

Which material to select for different types of fractures.


Biocompatibility and mechanical properties of materials.
What composition materials are needed.
Take a quite long time to check degradation behavior.

Machine learning is a revolutionary application that can be applied in


different ways in different industries. In this chapter, machine learning is
applied for the prediction of the bone healing process after the occurrence
of fractures in different parts of the body. For this, biodegradable material
selection and its degradation rate are very important factors that need to be
predicted and optimized for fast healing.

Some of the properties need to be studied while selecting the materials for
manufacturing of implant devices.

Modulus of elasticity: Elasticity must be comparable to natural bone, that


is, ~20GPa to for uniform stress distribution at implant for reduced rela-
tive movement at implant–​bone interface.

Tensile strength: Implants should have high tensile strength, which lowers
the stress at fracture interface.

Yield strength: Implants should have high yield strength to prevent brittle
fracture under cyclic load.

Elongation rate: It is the measure for deformation in implants that


determines ductility.

Degradation rate: Degradation rate should be related with the healing rate.
Implant materials should be dependent on the length of time that it is
necessary to keep them in the body.

Cytotoxicity: Implants should not be toxic to cells or tissues.

In this chapter, an algorithm is proposed for the prediction of material prop-


erties and its healing process according to the application area. Deciding the
implant material and its composition is a very confusing task, such as which
material is perfect to take as base material and its composition. The researcher
278

278 Machine Learning Algorithms and Applications in Engineering

FIGURE 16.3
Proposed training process.

has to make samples and perform testing on them to know their mechanical
and biological behavior. It is quite a long task to analyze the material. In any
situation, if the material testing fails, the entire effort becomes worthless.
So, to reduce time consumption, the machine learning approach is used to
predict the accuracy of material composition used for implant design. This
process is performed in two steps: training and testing. Figure 16.3 illustrates
the training process of the proposed model. Similarly, Figure 16.4 represents
the testing process of the model.

Algorithm: Degradation of Behavior Prediction


Initialization
Select implant location
Initialize the mechanical and biological properties
279

Machine Learning Application to Predict the Degradation Rate 279

FIGURE 16.4
Testing process.

Form S-​groups
Main Loop
While iter < max iteration
Select features
Perform training process
Aggregate training rules
Update decision
Final decision
End

In ensemble machine learning training, each random forest module is trained


independently with training samples and the data samples of each random
forest are correctly classified. Similarly, all other data values are trained inde-
pendently on individual random forest module and finally integrated as an
ensemble or a combination of several random forest modules, which will
expand the correctly classified area incrementally. This proposed ensemble
random forest training module performs better in case of prediction problems.
During the training phase of the proposed ensemble machine learning archi-
tecture, each random forest module is trained individually with training
samples from the dataset. This makes each trained random forest module
280

280 Machine Learning Algorithms and Applications in Engineering

be different from each other. Each random forest module can be trained with
different parameters of implant materials. Bagging, random selection, and
boosting selection strategies can be used to select training samples
In this proposed architecture, bagging rules are taken as a base for ensemble
module in which each module is trained individually and further they are
aggregated by applying a combination method. During the testing phase, the
aggregate strategy or voting strategy among all machine learning modules
will decide the test data class label. In ensemble random forest architecture,
n training sample sets are constructed with n individual modules. To achieve
higher efficiency, different training sample sets are taken in order to improve
the aggregation result with higher efficiency.

16.4 Conclusion
Biomedical implants are used in many applications such as hip replacement,
femur bone replacement, and dental implants. Research has been ongoing on
making better implant designs and selecting the best material composition
for making implants last longer, in addition to reducing the body’s reaction
to their presence. For proper healing of the bone, implant devices are required
to provide mechanical strength and stability. Choosing the proper material
for making implant devices can improve osseointegration. Mechanical
strength or tensile strength of the implants must be compatible with the nat-
ural bone. But selection of the material manually and its experimental ana-
lysis to determine its degradation rate take a very long time. The primary
focus of this work is to use the power of machine learning for effective bone
healing without any adverse effects. The use of machine learning will help in
deciding the suitable implant material to enable fast healing of the bone frac-
ture without any side effects.

References
[1] Claes, L. and Ignatius, A. Development of new, biodegradable implants. Chirurg.
73, 990–​996 (2002). https://​doi.org/​10.1007/​S00​104-​002-​0543-​0.
[2] Li, C., Guo, C., Fitzpatrick, V., Ibrahim, A., Zwierstra, M.J., Hanna, P., Lechtig, A.,
Nazarian, A., Lin, S.J., and Kaplan, D.L. Design of biodegradable, implantable
devices towards clinical translation. Nat. Rev. Mater. 51(5), 61–​81 (2019). https://​
doi.org/​10.1038/​s41​578-​019-​0150-​z.
[3] Hofmann, G.O. Biodegradable implants in traumatology: A review on the state-​
of-​the-​art. Arch. Orthop. Trauma Surg. 1143 (114), 123–​132 (1995). https://​doi.
org/​10.1007/​BF0​0443​385.
281

Machine Learning Application to Predict the Degradation Rate 281

[4] Liu, Y., Zheng, Y., and Hayes, B. Degradable, absorbable or resorbable—​what
is the best grammatical modifier for an implant that is eventually absorbed by
the body? Sci. China Mater. 60, 377–​391 (2017). https://​doi.org/​10.1007/​S40​
843-​017-​9023-​9.
[5] Karpouzos, A., Diamantis, E., Farmaki, P., Savvanis, S., and Troupis, T.
Nutritional aspects of bone health and fracture healing. J. Osteoporos. (2017).
https://​doi.org/​10.1155/​2017/​4218​472.
[6] Radha, R. and Sreekanth, D. Insight of magnesium alloys and composites for
orthopedic implant applications—​A review. J. Magnes. Alloy. 5, 286–​312 (2017).
https://​doi.org/​10.1016/​J.JMA.2017.08.003.
[7] Wang, W., Han, J., Yang, X., Li, M., Wan, P., Tan, L., Zhang, Y., and Yang, K. Novel
biocompatible magnesium alloys design with nutrient alloying elements Si, Ca
and Sr: structure and properties characterization. Mater. Sci. Eng. B. 214, 26–​36
(2016). https://​doi.org/​10.1016/​J.MSEB.2016.08.005.
[8] Li, H., Yang, H., Zheng, Y., Zhou, F., Qiu, K., and Wang, X. Design and
characterizations of novel biodegradable ternary Zn-​ based alloys with IIA
nutrient alloying elements Mg, Ca and Sr. Mater. Des. 83, 95–​102 (2015). https://​
doi.org/​10.1016/​J.MAT​DES.2015.05.089.
[9] Dehestani, M., Adolfsson, E., and Stanciu, L.A. Mechanical properties and
corrosion behavior of powder metallurgy iron-​hydroxyapatite composites for
biodegradable implant applications. Mater. Des. 109, 556–​569 (2016). https://​
doi.org/​10.1016/​J.MAT​DES.2016.07.092.
[10] Kottuparambil, R.R., Bontha, S., Rangarasaiah, R.M., Arya, S.B., Jana, A., Das,
M., Balla, V.K., Amrithalingam, S., and Prabhu, T.R. Effect of zinc and rare-​earth
element addition on mechanical, corrosion, and biological properties of magne-
sium. J. Mater. Res. 33, 3466–​3478 (2018). https://​doi.org/​10.1557/​JMR.2018.311.
[11] Drevet, R., Zhukova, Y., Malikova, P., Dubinskiy, S., Korotitskiy, A., Pustov, Y.,
and Prokoshkin, S. Martensitic transformations and mechanical and corrosion
properties of Fe-​Mn-​Si alloys for biodegradable medical implants. MMTA. 49,
1006–​1013 (2018). https://​doi.org/​10.1007/​S11​661-​017-​4458-​2.
[12] Tong, X., Zhang, D., Zhang, X., Su, Y., Shi, Z., Wang, K., Lin, J., Li, Y., Lin, J., and
Wen, C. Microstructure, mechanical properties, biocompatibility, and in vitro
corrosion and degradation behavior of a new Zn–​5Ge alloy for biodegradable
implant materials. Acta Biomater. 82, 197–​204 (2018). https://​doi.org/​10.1016/​
J.ACT​BIO.2018.10.015.
[13] Wątroba, M., Bednarczyk, W., Kawałko, J., Mech, K., Marciszko, M., Boelter,
G., Banzhaf, M., and Bała, P. Design of novel Zn-​Ag-​Zr alloy with enhanced
strength as a potential biodegradable implant material. Mater. Des. 183, 108154
(2019). https://​doi.org/​10.1016/​J.MAT​DES.2019.108​154.
[14] Xia, D., Liu, Y., Wang, S., Zeng, R.C., Liu, Y., Zheng, Y., and Zhou, Y. In vitro
and in vivo investigation on biodegradable Mg-​Li-​Ca alloys for bone implant
application. Sci. China Mater. 62, 256–​272 (2018). https://​doi.org/​10.1007/​S40​
843-​018-​9293-​8.
[15] Suryavanshi, A., Khanna, K., Sindhu, K.R., Bellare, J., and Srivastava, R.
Development of bone screw using novel biodegradable composite orthopedic
biomaterial: from material design to in vitro biomechanical and in vivo biocom-
patibility evaluation. Biomed. Mater. 14 (2019). https://​doi.org/​10.1088/​1748-​
605X/​AB1​6BE.
282

282 Machine Learning Algorithms and Applications in Engineering

[16] Zhang, Z.Y., Guo, Y.H., Zhao, Y.T., Chen, G., Wu, J.L., and Liu, M.P. Effect of
reinforcement spatial distribution on mechanical properties of MgO/​ ZK60
nanocomposites by powder metallurgy. Mater. Charact. 150, 229–​235 (2019).
https://​doi.org/​10.1016/​J.MATC​HAR.2019.02.024.
[17] Yang, H., Jia, B., Zhang, Z., Qu, X., Li, G., Lin, W., Zhu, D., Dai, K., and Zheng,
Y. Alloying design of biodegradable zinc as promising bone implants for load-​
bearing applications. Nat. Commun. 11, 1–​16 (2020). https://​doi.org/​10.1038/​
s41​467-​019-​14153-​7.
[18] Razzaghi, M., Kasiri-​ Asgarani, M., Bakhsheshi-​ Rad, H.R., and Ghayour, H.
Microstructure, mechanical properties, and in-​vitro biocompatibility of nano-​
NiTi reinforced Mg–​ 3Zn-​0.5Ag alloy: Prepared by mechanical alloying for
implant applications. Compos. Part B Eng. 190, 107947 (2020). https://​doi.org/​
10.1016/​J.COMP​OSIT​ESB.2020.107​947.
[19] Cilla, M., Borgiani, E., Martínez, J., Duda, G.N., and Checa, S. Machine learning
techniques for the optimization of joint replacements: Application to a short-​
stem hip implant. PLoS One. 12 (2017). https://​doi.org/​10.1371/​JOUR​NAL.
PONE.0183​755.
[20] Chatterjee, S., Dey, S., Majumder, S., RoyChowdhury, A., and Datta, S.
Computational intelligence based design of implant for varying bone conditions.
Int. J. Numer. Method. Biomed. Eng. 35 (2019). https://​doi.org/​10.1002/​
CNM.3191.
[21] Niculescu, B., Faur, C.I., Tataru, T., Diaconu, B.M., and Cruceru, M. Investigation
of biomechanical characteristics of orthopedic implants for tibial plateau
fractures by means of deep learning and support vector machine classification.
Appl. Sci. 10, 4697 (2020). https://​doi.org/​10.3390/​APP1​0144​697.
[22] Borjali, A., Monson, K., and Raeymaekers, B. Predicting the polyethylene wear
rate in pin-​on-​disc experiments in the context of prosthetic hip implants: Deriving
a data-​driven model using machine learning methods. Tribol. Int. 133, 101–​110
(2019). https://​doi.org/​10.1016/​J.TRIBO​INT.2019.01.014.
[23] Bedi P., Goyal S.B., Rajawat A.S., Shaw R.N., and Ghosh A. (2022) A framework
for personalizing atypical web search sessions with concept-​based user profiles
using selective machine learning techniques. In: Bianchini M., Piuri V., Das S.,
and Shaw R.N. (eds), Advanced Computing and Intelligent Technologies. Lecture
Notes in Networks and Systems, vol. 218. Springer, Singapore. https://​doi.org/​
10.1007/​978-​981-​16-​2164-​2_​23
newgenprepdf

283

17
Predicting the Outcomes of Myocardial
Infarction Using Neural Decision Forest

Akashdeep Singh Chaudhary and Ravinder Saini


Department of Computer Science and Engineering, Chandigarh University
Mohali, Punjab, India
[email protected]; [email protected]

CONTENTS
17.1 Introduction.............................................................................................. 284
17.2 Neural Decision Forests.......................................................................... 284
17.3 Literature Review..................................................................................... 286
17.4 Research Gap and Objective of the Research....................................... 289
17.5 Data Collection......................................................................................... 289
17.6 Data Analysis............................................................................................ 289
17.6.1 Attribute 113 as FIBR_PREDS................................................... 290
17.6.2 Attribute 114 as PREDS_TAH................................................... 290
17.6.3 Attribute 115 as JELUD_TAH................................................... 290
17.6.4 Attribute 116 as FIBR_JELUD................................................... 291
17.6.5 Attribute- 117 as A_V_BLOK.................................................... 291
17.6.6 Attribute 118 as OTEK_LANC................................................. 291
17.6.7 Attribute 119 as RAZRIV........................................................... 291
17.6.8 Attribute 120 as DRESSLER...................................................... 291
17.6.9 Attribute 121 as ZSN.................................................................. 291
17.6.10  Attribute 122 as REC_IM........................................................... 292
17.6.11 Attribute 123 as P_IM_STEN.................................................... 292
17.6.12  Attribute 124 as LET_IS............................................................. 292
17.7 Model Training......................................................................................... 292
17.7.1 Parameters Used for Training the Model................................ 292
17.7.2 Parameters for Training a Neural Decision Tree Model....... 293
17.7.3 Parameters for Training a Neural Decision
Forest Model................................................................................ 293
17.8 Results........................................................................................................ 293
17.9 Conclusions............................................................................................... 294

DOI: 10.1201/9781003104858-17 283


284

284 Machine Learning Algorithms and Applications in Engineering

17.1 Introduction
Myocardial infarction (MI) has become a serious silent killer in current times.
It is very hard to predict the likely outcome for the patient suffering from
MI. There are many factors involved such as diabetes, excessive alcohol
consumption, high blood pressure, lack of exercise, high blood cholesterol,
smoking, and lack of proper diet. Most of the patients suffering from MI have
coronary artery disease (CAD). MI is usually caused due to the coronary
artery blockage, which results from a breakage of an atherosclerotic plaque.
So, an early stage-​diagnosis with the help of tests such as electrocardiograms
(ECGs), blood tests, and coronary angiography may help locate the presence
of any disease. In general, chest pain is one of the symptoms that may occur
due to MI and may travel to the shoulder, to the arm, and may even to the
back. Feeling fainted, shortening of breath, cold sweat, or tiredness are a few
other symptoms that can be observed in the patient’s body. MI may lead to
cardiac arrest, failure of the heart, or even an abnormal heartbeat.
Predicting MI outcomes is quite challenging because a many factors affect
them. There is high mortality in the first year of patients suffering from
acute myocardial infarction (AMI). A large number of people die from AMI
even before reaching the hospitals. This is because of the huge uncertainty
in the prediction of MI complications and outcomes. It can occur without
complications or with complications. At the same time, it is observed that half
the number of patients having acute or subacute MI have complications that
result in worsening of the disease and even result in loss of life of the patient.
It is hard even for experienced specialists to predict the complications. But
with the help of a few techniques such as deep learning and using the pre-
vious data of the patients, it is quite feasible to some extent.

17.2 Neural Decision Forests


Neural decision forests (NDFs) are based upon two approaches. The first one
is convolution neural networks (CNNs) and the second is random forests. To
analyze visual imagery, the most commonly used artificial neural network is
CNN. CNN works on the principle of feature maps, which are nothing but
the translation equivariant responses produced by the filters that slide along
input features. It corresponds to a shared-​weight architecture of convolution
kernels. CNNs are used in the field of computer vision, video processing,
image processing, and natural language processing. The second approach
consists of random forest, which is the method for classification-​regression
and many more, which is an ensemble learning technique. The random
forest is the combination of interconnected trees, each of which is going to
285

Predicting the Outcomes of Myocardial Infarction 285

FIGURE 17.1
Architecture of the neural decision tree.

make predictions for the dataset provided as the input to it. Random forests
or random decision forests create a large number of decision trees at the
training time. In case of classification tasks, the class selected by most of
the trees is the output of the random forest. Whereas in case of regression,
the average prediction of the individual tree is the output. The purpose
of using the concept of random forest in the NDF is to reduce overfitting
to the training dataset. The deep neural decision forest (deep NDF) technique
becomes the bridge between classification trees and the representation
learning approach by training them in an end-​to-​end manner. Here, the
concept of differentiable and the stochastic decision tree model comes into
play. It propagates the representation learning organized in the initial layers
of a (deep) CNN.
The deep NDF is the large number of interconnected decision trees, each
of which makes a prediction for the given sample input and the average of
all these predictions is made to get the random forest prediction. Figure 17.1
shows the architecture of a single neural decision tree. In other words, it can
be said that random first uses a divide and conquer policy and has a simple
model with very high performance.
In the NDF model, there are a number of neural decision trees that are
trained at the same time. The average output of the trees is calculated as the
final output of the NDF.
NDF consists of many neural decision trees in which each tree has to learn
two types of weights, “pi” and “decision_​fn.” ‘Pi’ represents the probability
distribution of the classes that are present in the tree leaves, whereas “deci-
sion_​fn” represents the probability of going to each leaf node. There are
four steps in the working of the neural decision tree. In the first step, the
model takes the input features in the form of a single vector containing all
the features of an instance in the batch. This vector is generated with the
help of CNN. In the second step, the model randomly selects a subset of
286

286 Machine Learning Algorithms and Applications in Engineering

input features using “used_​features_​mask.” In the third step, for each input
instance taken from the second step, the model computes the probabilities
(mu) to reach the tree leaves. The model iteratively performs a stochastic
routing throughout the tree levels. Finally, in the fourth step, we get the final
output by combining the class probabilities at the leaves with the probabil-
ities of reaching the leaves.

17.3 Literature Review
Ibrahim el. al. [1]‌states that there is a need for accurate detection of AMI at
an early stage so that the patient can get timely provision of medical inter-
vention and is crucial for reducing the mortality rate. It is already known that
machine learning has proved its potential in aiding the diagnosis of diseases.
Lujain Ibrahim et al. [1] has used 713,447 extracted ECG samples along with
the related auxiliary data from the longitudinal and comprehensive ECG-​
ViEW II database for predicting the AMI. The author has also conducted
research using the XGBoost, which is a type of decision-​tree-​based model.
The research revolves around creating a framework that can be used to detect
AMI at an early stage. For the best-​performing CNN, recurrent neural net-
work (RNN) and XGBoost models, the prediction accuracy comes out to be
89.9, 84.6, and 97.5 percent, respectively, and ROC as 90.7, 82.9, and 96.5 per-
cent (the curve areas) are achieved, respectively. The importance and use of
the machine learning techniques are clearly proven to be of great value in the
prediction of cardiovascular disease in the paper.
Lenselink et al. [2]‌studied the risk prediction models (RPMs) for CAD
with the help of variables related to CAD. But the use of predictors in the
clinical practice is very limited because of unavailability of a proper descrip-
tion of the model, method for external validation, and the head-​to-​head
comparison. The author uses Tufts PACE CPM Registry and a systematic
PubMed search to identify the RPMs for CAD prediction and all the selected
models are externally verified in three large cohorts, namely, UK Biobank,
LifeLines, and PREVEND. The author takes two endpoints, which are MI
as a primary endpoint and CAD as a secondary endpoint, into consider-
ation for validating every RPM externally. It consists of MI, coronary artery
bypass grafting, and percutaneous coronary intervention. He calculates C-​
index (model discrimination), intercept and regression slope (calibration),
and accuracy (Brier Score) to compare the selected RPMs. To estimate the
calibration ability of an RPM, he used the method of linear regression ana-
lysis. In the paper, 28 RPMs were selected, but according to this paper, no
best-​performing RPM is identified as C-​index of most of the RPMs is 0.706
± 0.049, 0.778 ± 0.097, and 0.729 ± 0.074 for the prediction of MI in different
287

Predicting the Outcomes of Myocardial Infarction 287

cohorts of UK Biobank, LifeLines, and PREVEND, respectively. Therefore,


research concludes that no particular RPM could be suggested for the pre-
diction of risk of CADs.
Smith et al. [3]‌discusses various RPMs and carried out extensive research
on AMI. He compares the published literature related to RPM for the pre-
diction of AMI through March 2017 for 30-​day hospital readmission among
adults. He carried out research using 11 different studies on 18 unique RPMs
having varied settings basically in the United States. Sixteen models out of
18 unique RPMs are specific for AMI. Out of 18 RPMs, six models chosen
are based upon administrative data, four models are based upon electronic
health record data, three models are based upon clinical hospital data, and
the last five models are based upon the cardiac registry data. The range of the
median overall observed is 10.6–​21.0 percent (average 16.3 percent) across
studies conducted. The most frequently used predictors in the models are
demographics, comorbidities, and utilization metrics ranging from 7 to 37
in numbers. It is observed that the models have median C statistics as 0.65,
having range 0.53–​0.79. It is concluded from the studies that the models for
the prediction of AMI have limited predictive ability and do not have cer-
tain generalizability because of the methodological limitations. Also, it can
be said that no model provides the information related to the identification of
AMI, based on which suitable action can be taken at an early stage and risk-​
stratification of patients suffering from AMI can be done before discharging
from the hospital.
Mechanic et al. [4]‌discusses in detail about the acute MI and its results
such as irreparable damage to the heart muscles because of an insufficient
supply of oxygen. The patient may become prone to arrhythmias due to
impairment in systolic function and diastolic because of MI. There can be
a number of serious complications because of MI. Information related to
many factors such as smoking, dyslipidemia, obesity, hypertension, dia-
betes mellitus, poor oral hygiene, and sedentary lifestyle. which affect MI
is given in the paper. The paper discusses other causes of MI also which
included trauma, vasculitis, drug use (cocaine), coronary artery anomalies,
coronary artery emboli, aortic dissection, and excess demand on the heart
(hyperthyroidism, anemia). According to this research, at least one-​third of
the patients die before reaching the hospital and another 40–​50 percent
of patients have a fatal outcome upon reaching the hospital. Another 5–​10%
of the patients would face the same outcome within the first year after having
MI. The author makes the conclusion in the research that the physical exam-
ination of the patient must include the information related to vital signs,
appearance of the patient, diaphoresis, and lung findings along with cardiac
auscultation because it plays an important role to reveal more information.
Ventricular arrhythmia, atrial fibrillation, or tachycardia may be revealed by
observing the heart rate. The physical examination of the patient can reveal
various other crucial points, such as observing blood pressure of patients
288

288 Machine Learning Algorithms and Applications in Engineering

can reveal hypotension, tachypnea and fever can be common, and because
of the distended neck veins, there can be the right ventricular failure. Last,
but not the least, if the patient has developed pulmonary edema, then he
may have wheezing and rales.
Panju et al. [5]‌focuses on the features that help in increasing or decreasing
the probability of AMI. The selected features are history, physical examin-
ation of the patient, and ECG data. The ECG data is included for research
as the doctor usually interprets the results as the immediate initial clinical
assessment of the condition of the patient. In the first step, three diagnostic
groupings of the patients were made on the basis of acute chest pain and
then compared the contrast with the categorization of chest pain whether
the MI is present or not. The symptoms of MI are described briefly along
with the signs of MI mechanism of chest pain and the conditions that are
usually present with other symptoms related to MI. The paper discusses, in
detail, the role of the accuracy and precision of history, the physical exam-
ination of the patient, and ECG data in the identification of MI. The clinical
data related to these mentioned features along with their associated like-
lihood ratios (LRs) is considered for the prediction rules for AMI and are
taken with a broad set of inclusion criteria. The paper provides the conclu-
sion that the most crucial clinical feature is chest pain radiating toward the
arm that increases the probability of MI. Precisely, patients having MI have
two times the probability of having the chest pain radiating toward the left
arm than the patients who are not having MI. Meanwhile, the probability
of chest pain radiating toward the right arms is three times and the prob-
ability of chest pain radiating toward both the arms is seven times higher
than the patients without MI.
Rossiev et al. [6]‌discusses the computer expert system to forecast four
different types of complications that may appear to the patient suffering
from the MI in the hospital period. The neural network used in the paper
gathers the experience while training on the input data of real clinical cases.
The dataset chosen contains four types of attributes that have the high pri-
ority or are more crucial. They divide the main task into eight subtasks,
which contain four binary subtasks with the outcomes as 1 or 0 for having
the complication or not having the complications, respectively, and four
subtasks have the numerical output. For network training, backpropagation
technique is used. The objective of training is to minimize the estimation
function (global minimum taken). The conclusion made in this paper is
about the great possibilities of creating a neural network expert system for
predicting the complications of MI with the use of neural networks and
accelerating the process of creating an expert system that does not require
the mathematical algorithms for solving each task. The output from this
expert system is probabilistic. The doctor can take help from the expert
system by feeding the input to the expert system and make his or her own
decision afterward.
289

Predicting the Outcomes of Myocardial Infarction 289

17.4 Research Gap and Objective of the Research


From the literature review, it is found that less work has been done by taking
the myocardial infarction dataset, which contains the 111 number of input
parameters and with 12 outcome attributes (complications). It is very hard
to establish the relationship between so many attributes and then make the
prediction for the outcomes. The methods, techniques, approaches, and
algorithms discussed in the literature review do not deal with the missing
values in the dataset and thus give low accuracy and performance when
dealing with the missing values. Also, the algorithms generally take longer
to train the model, which is not good.
The main objective of this research is to make a model that solves the above
three problems. The model should be able to work on a large number of
input attributes as well as the model should not perform low on the dataset
that contains missing values. The model should take less time to train on the
dataset and with the least number of epochs, the model should give high
performance.

17.5 Data Collection
MI complications dataset are taken from UCI Machine Learning [7]‌. The
dataset has 1,700 instances with 124 unique attributes, but it has a few missing
values also (about 7.6 percent). Figure 17.2 shows the first 20 instances of the
dataset with the first header as the name of the attributes each starting with
a unique ID.
Out of 124 attributes, the first column is specified for the patient’s ID,
columns 2–​112 attributes are used as input data for prediction, and columns
113–​124 are attributes to be predicted that are the possible outcomes
(complications). The prediction is done for the time at the end of 72 h (72 h
after admission to the hospital) and columns (2–​112) are used as the input for
prediction.

17.6 Data Analysis
As there are a lot of attributes used in the dataset, we have discussed only the
attributes that are to be predicted (complications). We have given the details
regarding the values of the attributes, their significance in relation to the MI,
290

290 Machine Learning Algorithms and Applications in Engineering

FIGURE 17.2
First 20 instances of the dataset.

the number of instances, and their percentage present in the dataset. In the
dataset, the missing values are not present in the attributes that are to be
predicted (complications), but these are present in other remaining attributes
that are taken as input to the model. A brief description of the 12 attributes
(113–​124) used in the dataset, which are to be predicted, is given below.

17.6.1 Attribute 113 as FIBR_​PREDS


It stands for atrial fibrillation having the values “0” and “1.” Here, “0” stands
for “no” and constitutes for 1,530 instances (90.00%), whereas “1” stands for
“yes” and constitutes for 170 instances (10.00%) with no missing values.

17.6.2 Attribute 114 as PREDS_​TAH


It stands for supraventricular tachycardia having the values “0” and “1.”
Here, “0” stands for “no” and constitutes for 1,680 instances (98.82 percent),
whereas “1” stands for “yes” and constitutes for 20 instances (1.18 percent)
with no missing values.

17.6.3 Attribute 115 as JELUD_​TAH


It stands for ventricular tachycardia having the values “0” and “1.” Here, “0”
stands for “no” and constitutes for 1,658 instances (97.53 percent), whereas
291

Predicting the Outcomes of Myocardial Infarction 291

“1” stands for “yes” and constitutes for 42 instances (2.47 percent) with no
missing values.

17.6.4 Attribute 116 as FIBR_​JELUD


It stands for ventricular fibrillation having the values “0” and “1.” Here, “0”
stands for “no” and constitutes for 1,629 instances (95.82 percent), whereas
“1” stands for “yes” and constitutes for 71 instances (4.18 percent) with no
missing values.

17.6.5 Attribute-​ 117 as A_​V_​BLOK


It stands for the third-​degree AV block having the values “0” and “1.” Here,
“0” stands for “no” and constitutes for 1,643 instances (96.65 percent),
whereas “1” stands for “yes” and constitutes for 57 instances (3.35 percent)
with no missing values.

17.6.6 Attribute 118 as OTEK_​LANC


It stands for the third-​degree pulmonary edema having the values “0” and
“1.” Here “0” stands for “no” and constitutes for 1,541 instances (90.65 per-
cent), whereas “1” stands for “yes” and constitutes for 159 instances (9.35 per-
cent) with no missing values.

17.6.7 Attribute 119 as RAZRIV


It stands for myocardial rupture having the values “0” and “1.” Here, “0”
stands for “no” and constitutes for 1,646 instances (96.82 percent), whereas
“1” stands for “yes” and constitutes for 54 instances (3.18 percent) with no
missing values.

17.6.8 Attribute 120 as DRESSLER


It stands for Dressler syndrome having the values “0” and “1.” Here, “0”
stands for “no” and constitutes for 1,625 instances (95.59 percent), whereas
“1” stands for “yes” and constitutes for 75 instances (4.41 percent) with no
missing values.

17.6.9 Attribute 121 as ZSN


It stands for chronic heart failure having the values “0” and “1.” Here, “0”
stands for “no” and constitutes for 1,306 instances (76.82 percent), whereas
“1” stands for “yes” and constitutes for 394 instances (23.18 percent) with no
missing values.
292

292 Machine Learning Algorithms and Applications in Engineering

17.6.10 Attribute 122 as REC_​IM


It stands for relapse of the MI having the values “0” and “1.” Here, “0”
stands for “no” and constitutes for 1,541 instances (90.65 percent), whereas
“1” stands for “yes” and constitutes for 159 instances (9.35 percent) with no
missing values.

17.6.11 Attribute 123 as P_​IM_​STEN


It stands for post-​infarction angina having the values “0” and “1.” Here, “0”
stands for “no” and constitutes for 1,552 instances (91.29 percent), whereas
“1” stands for “yes” and constitutes for 148 instances (8.71 percent) with no
missing valu

17.6.12 Attribute 124 as LET_​IS


It stands for lethal outcome having the values “0,” “1,” “2,” “3,” “4,” “5”, “6,”
and “7.” This attribute has eight possible values. Here, “0” stands for “alive”
and constitutes for 1,429 instances (84.06 percent), “1” stands for “cardiogenic
shock” and constitutes for 110 instances (6.47 percent), “2” stands for “pul-
monary edema” and constitutes for 18 instances (1.06 percent), “3” stands
for “myocardial rupture” and constitutes for 54 instances (3.18 percent),
“4” stands for “progress of congestive heart failure” and constitutes for 23
instances (1.35 percent), “5” stands for “thromboembolism” and constitutes
for 12 instances (0.71 percent), “6” stands for “asystole” and constitutes
for 27 instances (1.59 percent), “7” stands for “ventricular fibrillation” and
constitutes for 27 instances (1.59 percent), with no missing values.

17.7 Model Training
For training the model, we have used 80–​20 split proportion for training and
testing, respectively. The attributes of columns 113–​124 are taken individually
with the dataset of columns 1–​112 for the prediction separately. We calculated
the performance of the model at different instances by varying different
parameters such as the number of epochs, the number of trees, and depth of
trees. The objective of research is to find the best NDF model, which gives the
highest performance for predicting the attributes 113–​124 (complications) by
varying different parameters that are discussed below.

17.7.1 Parameters Used for Training the Model


batch_​size: 100
num_​epoch: 10
293

Predicting the Outcomes of Myocardial Infarction 293

17.7.2 Parameters for Training a Neural Decision Tree Model


depth: 5
used_​feature_​rate: 1.0

17.7.3 Parameters for Training a Neural Decision Forest Model


num_​trees: 15
depth: 5
used_​feature_​rate: 0.50

By changing the value of the “used_​features_​rate” variable, we can change


the number of features to be used by each tree. For training a single neural
decision tree, the depth of the tree used is 5 and all the input features are
selected as the input to the model by defining used_​feature_​rate equal to
1. For training the NDF model, the number of the trees (num_​trees) is equal to
15, and the depth of each tree is equal to 5. We have randomly selected 50 per-
cent of the input features for input to every tree by defining used_​feature_​
rate equal to 0.5 while training the neural decision forest. The parameters
with these values give the best performance.

17.8 Results
We evaluate our model on the input data at various parameters. Basically, we
checked for the loss and sparse categorical accuracy of the model and found
the best suitable parameters for the given objective and the input data. The
best performance of the model is achieved by taking batch size as 100, and
the number of epochs as 10.
In the case of a single neural decision tree model, we use the depth as 5 and
we take all features for input features (used_​feature_​rate =​1.0).
In the case of neural decision forest, we use the number of trees as 10 (num_​
trees =​10) and the depth as 10, while taking only half the number of total
input features (used_​feature_​rate =​0.5), which are selected randomly.
The performance of a single neural decision tree model where we use all
input features columns 2–​112 for the prediction of attributes 113–​124 (stated
above) is shown in Table 17.1. The number of epochs used is 10, and the
depth of the tree used is 5.
The performance of the NDF that consists of many neural decision trees
where we use all input features columns 2–​112 for the prediction of attributes
113–​124 (stated above) is shown in Table 17.2. The number of epochs used is
10, number of trees used is 5, and the depth of trees used is 5.
294

294 Machine Learning Algorithms and Applications in Engineering

TABLE 17.1
Performance of a Single Neural Decision Tree Model

Attributes for Sparse categorical


S.No. the prediction Loss accuracy Test accuracy

1 FIBR_​PREDS 0.5234 0.8059 80.59%


2 PREDS_​TAH 0.1115 0.9824 98.24%
3 JELUD_​TAH 0.1659 0.9647 96.47%
4 FIBR_​JELUD 0.2706 0.9235 92.35%
5 A_​V_​BLOK 0.239 0.9353 93.53%
6 OTEK_​LANC 0.6142 0.6176 61.76%
7 RAZRIV 0.5007 0.8412 84.12%
8 DRESSLER 0.1405 0.9824 98.24%
9 ZSN 0.6586 0.6441 64.41%
10 REC_​IM 0.4775 0.8029 80.00%
11 P_​IM_​STEN 0.338 0.9 90.00%
12 LET_​IS 0.865 0.8324 83.24%

TABLE 17.2
Performance of the Neural Decision Forest Model

Attributes for Sparse categorical


S.No. the prediction Loss accuracy Test accuracy (%)

1 FIBR_​PREDS 0.5429 0.8324 83.24


2 PREDS_​TAH 0.1073 0.9824 98.24
3 JELUD_​TAH 0.1603 0.9647 96.47
4 FIBR_​JELUD 0.2711 0.9235 92.35
5 A_​V_​BLOK 0.2396 0.9353 93.53
6 OTEK_​LANC 0.4113 0.8559 85.59
7 RAZRIV 0.5033 0.8412 84.12
8 DRESSLER 0.1536 0.9824 98.24
9 ZSN 0.5476 0.7941 79.41
10 REC_​IM 0.4634 0.8441 84.41
11 P_​IM_​STEN 0.336 0.9 90.00%
12 LET_​IS 0.9169 0.7353 73.53%

17.9 Conclusions
The use of machine learning in the field of health care and medicine has
shown very good results and opens up tremendous possibilities in this field
for the benefit of the society. Machine learning also provides clues related to
diseases, which are quite difficult to predict with mathematical calculations
alone, by using sophisticated techniques such as NDFs. By the use of the
NDF method, the prediction of the possible outcomes (complications) of a
disease becomes easy, and it provides high accuracy despite having missing
295

Predicting the Outcomes of Myocardial Infarction 295

values in the input data. This gives the opportunity for patients to become
aware of the symptoms of the disease in advance and prepare accordingly or
get treatment as soon as possible.
In future work, the creation of more complex machine learning algorithms
is required to increase the efficiency of the prediction models. Moreover,
there is a need to create an algorithm that can work efficiently, not only for
the people of a particular geographical area, but for people who belong to
different communities, different places, and different geographical locations.
To avoid overfitting, the dataset should include the instances of the patients
having variety in terms of the demographics, lifestyle, age, and medical
history.

References
[1] Ibrahim L,Mesinovic M, Yang K-​W, Eid MA. Explainable prediction of acute
myocardial infarction using machine learning and Shapley values, IEEE Access,
2020;8: 210410–​210417, , doi: 10.1109/​ACCESS.2020.3040166.
[2] Lenselink C,Ties D,Pleijhuis R, van der Harst P. Validation and comparison of
28 risk prediction models for coronary artery disease, Eur. J. Prev. Cardiol., 2021.
doi: 10.1093/​eurjpc/​zwab095.
[3] Smith LN et al., Acute myocardial infarction readmission risk predic-
tion models, Circ. Cardiovasc. Qual. Outcomes, 2018; 11(1). doi: 10.1161/​
CIRCOUTCOMES.117.003885.
[4] Mechanic OJ, Gavin M, Grossman SA. Acute Myocardial Infarction. [Updated
2022 May 9]. In: StatPearls [Internet]. Treasure Island (FL): StatPearls Publishing;
2022 Jan. Available from: www.ncbi.nlm.nih.gov/​books/​NBK459​269/​.
[5] Panju AA, Is this patient having a myocardial infarction? JAMA, 1998;280(14):1256,
doi: 10.1001/​jama.280.14.1256.
[6] Rossiev DA, Golovenkin SE, Shulman VA, Matjushin GV. Neural networks for
forecasting of myocardial infarction complications, in The Second International
Symposium on Neuroinformatics and Neurocomputers, pp. 292–​ 298,
doi: 10.1109/​ISNINC.1995.480871.
[7] https:// ​ a rch ​ i ve.ics.uci.edu/ ​ m l/ ​ d atas ​ e ts/ ​ M yo ​ c ard ​ i al+​ i nf​ a rct​ i on+​ c ompli​
cati​ons
[8] Saleh M, Ambrose JA. Understanding myocardial infarction. F1000Res.
2018 Sep 3;7:F1000 Faculty Rev-​ 1378. doi: 10.12688/​f1000research.15096.1.
PMID: 30228871; PMCID: PMC6124376.
[9] Wu J, Qiu J, Xie E, Jiang W, Zhao R, Qiu J, Zafar MA, Huang Y, Yu AC. Predicting
in-​hospital rupture of type A aortic dissection using Random Forest, J. Thorac
Dis., 2019;11(11):4634–​4646.
[10] Chandler AB, Chapman I, Erhardt LR, et al. Coronary thrombosis in myocar-
dial infarction. Report of a workshop on the role of coronary thrombosis in the
pathogenesis of acute myocardial infarction. Am. J. Cardiol. 1974;34(7):823–​833.
doi: 10.1016/​0002-​9149(74)90703-​6
296

296 Machine Learning Algorithms and Applications in Engineering

[11] Thygesen K, Alpert JS, Jaffe AS, et al.: Third universal definition of myocardial
infarction. Circulation. 2012;126(16):2020–​35. doi: 10.1161/​CIR.0b013e31826e1058
[12] DeWood MA, Spores J, Notske R, et al. Prevalence of total coronary occlusion
during the early hours of transmural myocardial infarction. N Engl J Med.
1980;303(16):897–​902. doi: 10.1056/​NEJM198010163031601
[13] Ambrose JA, Najafi A. Strategies for the prevention of coronary artery disease
complications: Can we do better? Am J Med. 2018; pii: S0002–​9343(18)30382-​6.
doi: 10.1016/​j.amjmed.2018.04.006
[14] Bassand JP, Hamm CW, Ardissino D, et al. Guidelines for the diagnosis and
treatment of non-​ST-​segment elevation acute coronary syndromes: The task force
for the diagnosis and treatment of non-​ST-​segment elevation acute coronary
syndromes of the European Society of Cardiology. Eur Heart J 2007;28:1598–​660.
[15] Mandelzweig L, Battler A, Boyko V, et al. The second Euro Heart Survey
on acute coronary syndromes: Characteristics, treatment, and outcome of
patients with ACS in Europe and the Mediterranean Basin in 2004. Eur Heart J
2006;27:2285–​2293.
[16] Sanchis-​Gomar F, Perez-​Quilis C, Leischik R, Lucia A. Epidemiology of coronary
heart disease and acute coronary syndrome. Ann Transl Med 2016;4:256–​256.
[17] Frohlich ED, Quinlan PJ. Coronary heart disease risk factors: public impact of
initial and later-​announced risks. Ochsner J 2014;14:532–​537.
newgenprepdf

297

18
Image Classification Using Contrastive
Learning

Abhyuday Trivedi, Anjali Hembrom, Arkajit Saha, Tahreem Fatima,


Shreya Dey, Monideepa Roy, and Sujoy Datta
School of Computer Engineering, KIIT, Deemed University, Bhubaneswar, Odisha, India
[email protected]

CONTENTS
18.1 Introduction.............................................................................................. 298
18.2 Background............................................................................................... 298
18.3 Implementation........................................................................................ 300
18.3.1 Self-Supervised Learning.......................................................... 300
18.3.2 Contrastive Learning................................................................. 300
18.3.3 SimCLR........................................................................................ 301
18.3.3.1 Dataset......................................................................... 302
18.3.3.2 Data Augmentation................................................... 302
18.3.3.3 Extraction of Representation Vectors with
Neural Network Encoder......................................... 303
18.3.3.4 Nonlinear Projection Head....................................... 306
18.3.3.5 Normalized Temperature-Scaled
Cross-Entropy Loss Function for
Contrastive Prediction.............................................. 307
18.3.3.6 Training of ResNet-18 using NT- Xent Loss........... 308
18.4 Results and Discussion............................................................................ 309
18.4.1 Data Augmentation: Original and Augmented Image......... 309
18.4.2 Layers of Neural Network Base Encoder and
Projection Head.......................................................................... 310
18.4.3 Training Losses............................................................................311
18.4.4 Training of ResNet-18 using NT-Xent Loss............................ 312
18.5 Conclusion and Future Work................................................................. 312

DOI: 10.1201/9781003104858-18 297


298

298 Machine Learning Algorithms and Applications in Engineering

18.1 Introduction
The problem of learning visual representations effectively without human
supervision can be done by contrastive learning. This work has been based
on learning visual representations by using the SimCLR framework [1]‌.
Visual representations are image representation vectors on which supervised
linear or unsupervised classifiers can be trained for accurate image recogni-
tion and classification. These representations can be learned by training deep
learning models like ResNets on labeled datasets like ImageNet. But labeling
and annotating data is a time-​consuming and elaborate process and can be
avoided by using alternate learning techniques like self-​supervised learning
where the training data is automatically labeled by finding and utilizing
correlations between various input features.
Contrastive learning of visual representations can be done by efficiently
finding similar and dissimilar images. For understanding contrastive
representation learning, we interpreted the key elements of the SimCLR
framework which show that:

• A combination of various different data augmentation operations


is important for the contrastive prediction tasks to yield effective
representations and stronger data augmentation benefits unsupervised
contrastive learning more than supervised learning.
• Nonlinear transformation, which has the ability to learn between the
illustration and the contrastive loss, is introduced and considerably
improves the learned representations’ quality.
• Normalized embeddings and an adjusted temperature parameter posi-
tively affects representation learning with contrastive cross-​entropy loss.
• Larger batch sizes and longer training benefits contrastive learning
more compared to its supervised counterpart.

The rest of the chapter will describe the background of previous work done
in the field of using self-​supervised and contrastive learning to achieve state-​
of-​the-​art results; how SimCLR framework has been used and implemented
in this work; visualization of the results obtained from the implementation
and future scope and applications of our work.

18.2 Background
In this chapter, we explored and highlighted how a simple framework for
contrastive learning of visual representations referred to as SimCLR [1]‌
299

Image Classification Using Contrastive Learning 299

helps us to learn the visual representations. We have used ResNet-​18 and


trained our models for 20 epochs to get a particular convergence and show
how SimCLR benefits more from unsupervised contrastive learning than
supervised learning.
The idea of classifying representations of an image on the basis of similarity
with each other under small transformations dates long back. Dating back
to those approaches [2]‌, which used to learn representations by contrasting
positive pairs against negative pairs, SimCLR extended them by highlighting
and implementing recent advances in data augmentation, neural network
architecture, and contrastive loss. The individual components of SimCLR
have previously been researched upon, but here the improved and integrated
version, a result of a combination of different design choices, which provides
better performance has been used and described in the Results and Discussion
section of the chapter.
SimCLR has been used as the basis method of contrastive learning as some
ways by which SimCLR is better for visual representation than previously
proposed contrastive representation learning methods are:

• DIM/​AMDIM [3]‌[4] modified a ResNet and placed restrictions on


the receiving fields of the convolutional neural network (e.g., 3 ×
3 Convs were replaced with 1 × 1 Convs). In SimCLR framework,
the decoupled task, which has to be predicted, and neural network
base encoder architecture uses the final vector representations of
two augmented image vectors for prediction, thus needing powerful
ResNets.
• CPC v1 and v2 [5]‌[6] used a deterministic strategy to define the context
prediction task and a context aggregation network (for instance, a Pixel-​
CNN). SimCLR framework does not need such a context aggregation
network as it decouples the prediction task and the neural network base
encoder. In both these above pointers discussing different methods, it is
important to note that SimCLR uses simplified data augmentation and
that its NT-​Xent Loss function uses normalization and adjustment of
temperature parameter to give better similarity scores.
• InstDisc, MoCo, and PIRL [7]‌[8][9] generalized the Exemplar approach
originally proposed by [10] and leveraged an explicit memory bank.
SimCLR, on the other hand, doesn’t use a memory bank but in-​batch
negative example sampling is utilized.
• CMC [11] used a separate network for each view, whereas SimCLR only
shares only a single network among all the random augmented views.
• Similarity between augmented and un-​augmented copies of the same
original image are enhanced to the maximum possible value, but
SimCLR applies data augmentation fairly to both symmetric branches
of its framework [12].
300

300 Machine Learning Algorithms and Applications in Engineering

18.3 Implementation
18.3.1 Self-​Supervised Learning
Self-​
supervised learning method is used to train computers to do tasks
without the need to provide labeled data manually (Figure 18.1). It is a
subset of un-​supervised learning where results are derived by information
labeling, categorizing, and finally analyzing it. The conclusions are drawn by
the machine itself on the basis of connections and correlations. The system
learns and attains the capability to know the different parts of any object by
encoding it, so that recognition can be done from any angle. Only then can
the object be classified correctly and provide context for analysis to come up
with the desired output.

18.3.2 Contrastive Learning
Contrastive learning is a method used for finding similar and dissimilar
things by training a ML model to learn classification between similar and
dissimilar images (Figure 18.2). Contrastive learning contrasts positive pairs
against negative pairs and learns representations. This method of learning

FIGURE 18.1
Self-​supervised learning results.

FIGURE 18.2
Expected contrastive learning prediction output.
301

Image Classification Using Contrastive Learning 301

first learns generic representations of images from an unlabeled dataset and


fine-​tuned with a small dataset of labeled images for classification purposes.
Contrastive learning can be applied by

• generation of certain sized image batches,


• application of a transformation function to get a pair of two modified
images,
• passing the pair obtained through an encoder to get image
representations,
• passing the representations of the two augmented images through a
nonlinear dense layer, and
• getting an embedding vector for each augmented image in the batch.

The agreement between two augmented versions of an image is determined


using loss function’s cosine similarity feature. To find the probability of how
similar these two images are, a soft-​max function is applied.

18.3.3 SimCLR
SimCLR maximizes similarity value between differently augmented versions
of same image data by contrastive loss in latent space and learns visual
representations (Figure 18.3). The framework includes the following major
processes:

• To begin, SimCLR randomly draws examples from the original


dataset, transforms each example twice using a combination of simple
augmentations (random cropping, resize, and color distortion), creating

FIGURE 18.3
SimCLR framework.
302

302 Machine Learning Algorithms and Applications in Engineering

two sets of corresponding views. This comprises the data augmentation


module.
• SimCLR then computes the image representation vectors from the
augmented images using a neural network base encoder function and
for this purpose convolutional neural network (CNN) variant based on
the ResNet-​18 architecture is chosen.
• It maps a nonlinear projection of the image representation using a neural
network projection head of a fully connected network (i.e., MLP), which
amplifies the invariant features and maximizes the ability of identifica-
tion of different transformations of the same image.
• Then, a stochastic gradient descent is used to update both CNN and
MLP in order to minimize the loss function of the contrastive objective
for contrastive prediction task.

These components of the framework have been implemented and


highlighted in the following subsections.

18.3.3.1 Dataset
We have used a manually created ImageNet dataset containing 1250 images
for training (250 images for each of the 5 categories) and 250 images for
test (50 images for each of the five categories). The five categories used for
SimCLR framework analysis are car, dog, bear, donut, and jean.

18.3.3.2 Data Augmentation
Data augmentation has previously been used for supervised or unsupervised
representation learning [4,6,13] for contrastive prediction tasks by making
changes in the architecture. In SimCLR application, generally, for a batch of
N images, on applying the below mentioned composition of data augmenta-
tion operations, 2N augmented images are obtained as shown in Figure 18.4.
Then, for a particular positive pair of images (i, j) from the 2N images,
2(N–​1) images are considered as negative pair (i, j) examples as displayed in
Figure 18.5.
In our analysis, as ImageNet images are always of different sizes; each of
the images in our dataset went through the first data augmentation operation
and were cropped and resized to size 224 × 224 (Figure 18.6). This was done
by standard random cropping [14] of random size 1.0 in area, of original size
and then resizing of the cropped image. Thus, we obtained the first set of
augmented images. This implementation is carried out in Pytorch and makes
use of RandomResizedCrop class of torch package transforms of Pytorch
library as shown in the pseudocode in Figure 18.7. Additionally, in another
data augmentation operation, color distortion was done to get another set
303

Image Classification Using Contrastive Learning 303

FIGURE 18.4
2N images formed from batch of N images on performing data augmentation operations.

FIGURE 18.5
From the 2N images, 2(N–​1) images considered as negative pairs.

of augmented images where color distortion is composed of strong color


jittering controlled by using strength parameters and color dropping.

18.3.3.3 Extraction of Representation Vectors with Neural


Network Encoder
The next component of the framework feeds each positive pair obtained
from previous process in any neural network (standard ResNets) encoder
304

304 Machine Learning Algorithms and Applications in Engineering

FIGURE 18.6
Data augmentation highlighted in main SimCLR algorithm.

FIGURE 18.7
Data augmentation displayed in pseudo code.
305

Image Classification Using Contrastive Learning 305

FIGURE 18.8
Encoder function for representation vectors in main SimCLR algorithm.

FIGURE 18.9
ResNet-​18 with top layer replaced by fully connected layers and last layer replaced with
nonlinear classifier.

shown as f to obtain representation vectors represented as shown in


Figure 18.8. The representation vectors are the output obtained after the
average pooling layer.
We modified the SimCLR framework to compute the image representation
vectors from the augmented images using a neural network base encoder
function, but for this analysis, convolutional neural network (CNN) variant
based on the ResNet-​18 [15] architecture is chosen where the top layer is
replaced by some fully connected layers as shown in Figure 18.9.
306

306 Machine Learning Algorithms and Applications in Engineering

18.3.3.4 Nonlinear Projection Head


This component of the framework uses the representation vectors h obtained
from the standard encoding layer f and uses a nonlinear projection head g
(Figure 18.10). As the hidden layer before nonlinear projection head gives
a better representation than the layer after, the hidden layer is preferred for
representation. The vectors obtained as z are trained to remain unaltered
after data transformation. On using the projection head g, more information
is formed and retained at h; as the representation h is used before nonlinear
projection due to loss of more information at g because of contrastive loss [1]‌.
As already mentioned in the above component of the framework being
analyzed, we use ResNet-​18 as the base encoder network and a two-​layer
nonlinear MLP projection head is used to project the representation further.
Figure 18.11 shows how the last layer and ReLU activation layer is removed
incrementally and visualized using tSNE visualization while the Figure 18.12

FIGURE 18.10
Nonlinear projection head used after representation in main SimCLR algorithm.

FIGURE 18.11
Removal of last and ReLU layer and the method to see the tSNE visualization.
307

Image Classification Using Contrastive Learning 307

shows the removal of the last layer of the projection head and its affect can be
visualized using tSNE visualization.

18.3.3.5 Normalized Temperature-​Scaled Cross-​Entropy


Loss Function for Contrastive Prediction
Our motive is minimizing normalized temperature-​ scaled cross-​entropy
loss or NT-​Xent Loss [1]‌for positive pair of examples, which maximizes the
similarity between vectors zi and zj and at the same time makes the dissimi-
larity with all other vectors evident. Figure 18.13 shows the asymmetric loss
function used here.
For maximizing the agreement between image vectors zi and zj for analysis,
we used NT-​ Xent Loss optimized using SGD optimizer with learning
rate 1.2 (=​√Batch size * 0.075) and weight decay value10–​6 (Figure 18.14). We

FIGURE 18.12
Removal of last layer of projection head and the method to see the tSNE visualization.

FIGURE 18.13
Loss function for similarity maximization between vectors.
308

308 Machine Learning Algorithms and Applications in Engineering

FIGURE 18.14
NT-​Xent Loss function code.

FIGURE 18.15
Linear classifier code.

train a batch size 256 using cloud GPU and carry out the decaying of learning
rate with cosine decay schedule without restarts. We use SGD optimizer with
square root learning rate scaling for analysis purpose as we have considered a
small batch size and small number of epochs, but training with SGD becomes
unstable with large batch size; in that case, LARS optimizer can be used [1]‌.
Figure 18.14 shows the code for the loss function where tau is a temperature
hyperparameter, which makes loss function more expressible and here, the
similarity between two vectors a and b is the dot product of their respective
unit vectors a_​cap and b_​cap.

18.3.3.6 Training of ResNet-​18 using NT-​Xent Loss


A linear classifier is trained on the representations obtained from the second
last layer of projection head as the representation of this layer is better than
that of the one after it, whose code is visible in Figure 18.15. The training and
309

Image Classification Using Contrastive Learning 309

testing accuracy and losses versus number of epochs graph are generated
which is discussed upon in the next section.

18.4 Results and Discussion


In this chapter, we have analyzed how different components act on our
dataset to yield results, which helps us to draw a number of conclusions.
Here, we will highlight and discuss the results and related inferences.

18.4.1 Data Augmentation: Original and Augmented Image


We went ahead and applied a combination of two data augmentation
operations: first, random crop and resize and then color distortion as the
second; two times for each image, to get two new images. Following is the
result we obtained (Table 18.1).

TABLE 18.1
Two Images Each from “Donut” and “Bear” Category Obtained by Applying
Composition of Random Crop and Resize and Color Distortion on the Original Picture

Original image After applying composition of After applying composition of


random crop and Resize with random crop and resize with
Color distortion one time Color distortion two times
310

310 Machine Learning Algorithms and Applications in Engineering

The conclusion here is that even though the positive pairs are getting iden-
tified efficiently, but there exists no single transformation obtained after data
augmentation that is enough to learn good representations; but the above
discussed composition is good for learning generalizable features.

18.4.2 Layers of Neural Network Base Encoder and Projection Head


After the augmentation of images, we fed each positive pair in a convolu-
tional neural network having ResNet-​18 architecture and two-​layer projec-
tion head to project the representations further. Figure 18.11 showed how the
last and ReLU activation layers were removed while the Figure 18.12 showed
the removal of the last layer of the projection head. The t-​SNE visualization
of these above actions are shown in Tables 18.2 and 18.3.

t-​
SNE visualization of training dataset t-​
SNE visualization of testing
dataset

The conclusions drawn from the above discussion and results are as
follows:

• Removal of hidden layer and ReLU activation before projection head is


a better representation of the image vectors.
• Presence of clusters in above visualizations show that correct pairs of
images are being identified.

TABLE 18.2
t-​SNE Visualizations of the Last Layer Vectors of Train (10% of 1,250 =​125) and Test
(250) Images

t-​SNE visualization of training dataset t-​SNE visualization of testing dataset


311

Image Classification Using Contrastive Learning 311

TABLE 18.3
t-​SNE Visualizations of the Second Last Layer Vectors of Train (10% of 1250 =​125) and
Test (250) Images

t-​SNE visualization of training dataset t-​SNE visualization of testing dataset

FIGURE 18.16
Code to visualize training losses graph.

• Visualization of dataset on RHS of above tables are considered to be low


rank [1]‌as they are poorly separate as compared to the dataset on the
other side of the tables.

18.4.3 Training Losses
We modified the architecture of our Resnet-​18 model by replacing top and
last layers of it by fully connected layers and a classifier respectively. On
training the Resnet-​18 model, some training losses calculated by NT-​Xent
Loss function alongside the number of epochs used for training is visualized
using matplotlib.pyplot (Figures 18.16 and 18.17).
312

312 Machine Learning Algorithms and Applications in Engineering

FIGURE 18.17
Training losses versus no. of epochs graph.

18.4.4 Training of ResNet-​18 using NT-​Xent Loss


A linear classifier is trained on the representations obtained from the second
last layer of projection head as the representation of this layer is better than
that of the one after it, whose code is visible in Figure 18.15. The training
and testing accuracy and losses versus number of epochs (no. of epochs used
here=​20) graph are generated in Table 18.4.
On training of ResNet-​18 using NT-​Xent Loss, we obtain a test accuracy
of 64.8 percent, which is 12 percent lower than that obtained from training
ResNet-​18 using supervised classifier which gives accuracy of 76.8 percent.

18.5 Conclusion and Future Work


In this work, we present a simple framework for contrastive visual represen-
tation learning, which is an improved technology we have used here for
image recognition and classification. It is a high-​performance model with
high accuracy in which labeled samples are used for optimizing model
parameters.
313

Image Classification Using Contrastive Learning 313

TABLE 18.4
Accuracy and Losses Graphs Plotted While Training a Linear Classifier on 10 Percent
Labeled Training Data

We take a close look at its components and demonstrate the results of


various design choices. We build on previous approaches for self-​supervised,
semi-​supervised, and transfer learning and integrating our results and ana-
lysis with it. Data augmentation, usage of a nonlinear head at the neural
network’s end, and a normalized loss function distinguish our method from
normal supervised learning on ImageNet dataset.
With applications in facial recognition, driverless vehicles, medical dis-
ease detection, and even education, image recognition using SimCLR has
enhanced online visibility. Thus, image recognition systems have a bright
future.
Driverless vehicles, as well as amazing facial recognition mechanics, have
already been introduced in a number of countries, allowing for quicker and
more accurate object detection in real time. This was made possible by image
recognition using the SimCLR algorithm. Hence, it has become easy to expect
that the image recognition market will thrive in the upcoming years.

References
1. Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton,
“A simple framework for contrastive learning of visual representations”, In
the Proceedings of the 37th International Conference on Machine Learning,
Vienna, Austria, July 2020.
314

314 Machine Learning Algorithms and Applications in Engineering

2. Hadsell, R., Chopra, S., and LeCun, Y., “Dimensionality reduction by learning
an invariant mapping”, In 2006 IEEE Computer Society Conference on
Computer Vision and Pattern Recognition (CVPR’06), volume 2, pp. 1735–​
1742. IEEE, 2006.
3. R Devon Hjelm, Alex Fedorov, Samuel Lavoie-​Marchildon, Karan Grewal,
Phil Bachman, Adam Trischler, and Yoshua Bengio, “Learning deep
representations by mutual information estimation and maximization”, arXiv
preprint arXiv:1808.06670, August 2018.
4. Philip Bachman, R Devon Hjelm, and William Buchwalter, “Learning
representations by maximizing mutual information across views”, In the
33rd Conference on Neural Information Processing Systems (NeurIPS 2019),
Vancouver, Canada, 2019.
5. Aaron van den Oord, Yazhe Li, and Oriol Vinyals, “Representation learning
with contrastive predictive coding”, arXiv preprint arXiv:1807.03748, 2018.
6. Olivier J. Henaff, Aravind Srinivas, Jeffrey De Fauw, Ali Razavi, Carl Doersch,
S. M. Ali Eslami, and Aaron van den Oord, “Data-​efficient image recogni-
tion with contrastive predictive coding”, In the Proceedings of the 37th
International Conference on Machine Learning, Vienna, Austria, 2020
7. Zhirong Wu, Yuanjun Xiong, Stella Yu, and Dahua Lin, “Unsupervised fea-
ture learning via non-​parametric instance discrimination”, In Proceedings of
the IEEE Conference on Computer Vision and Pattern Recognition, pp. 3733–​
3742, 2018.
8. Kaiming He Haoqi Fan Yuxin Wu Saining Xie and Ross Girshick, “Momentum
contrast for unsupervised visual representation learning”, arXiv preprint
arXiv:1911.05722, 2019.
9. Ishan Misra and Laurens van der Maaten, “Self-​ supervised learning of
pretext-​invariant representations”, arXiv preprint arXiv:1912.01991, 2019.
10. Alexey Dosovitskiy, Philipp Fischer, Jost Tobias Springenberg, Martin
Riedmiller, and Thomas Brox, “Discriminative unsupervised feature learning
with exemplar convolutional neural networks”, In Advances in Neural
Information Processing Systems, pp. 766–​774, 2014.
11. Yonglong Tian, Dilip Krishnan, and Phillip Isola, “Contrastive Multiview
coding”, arXiv preprint arXiv:1906.05849, 2019.
12. Ye, M., Zhang, X., Yuen, P. C., and Chang, S.-​F. “Unsupervised embedding
learning via invariant and spreading instance feature”, In the Proceedings of
the IEEE Conference on Computer Vision and Pattern Recognition, pp. 6210–​
6219, 2019.
13. Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton, “Imagenet classi-
fication with deep convolutional neural networks”, In Advances in Neural
Information Processing Systems, pp. 1097–​1105, 2012.
14. Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed,
Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew
Rabinovich, “Going Deeper with Convolutions”, In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition, pp. 1–​9, 2015.
15. https://​med​ium.com/​analyt​ics-​vid​hya/​unders​tand​ing-​sim​clr-​a-​sim​ple-​
framew​ork-​for-​cont​rast​ive-​learn​ing-​of-​vis​ual-​repr​esen​tati​ons-​d544a​9003​f3c.
315

Index

Note: Page numbers in bold refer to tables and those in italic refer to figures.

accuracy 122 153; DWTSVM 149; Hilbert–​Haung


ACF plot 234 transform and support vector
activation synthesis theory 95–​6 machine (HHT-​SVM) 148; HMM and
acute myocardial infarction (AMI) 284 hierarchical clustering 152; novel
adapted convolutional neural network shilling attack detection 150–​1; PCA
(ADCNN) 75 and perturbation 149–​150; RDMA 150;
adaptive neuro fuzzy inference system re-​scale AdaBoost 148; SEMI-​SAD
(ANFIS) 21 150; statistical detection algorithm
additive manufacturing (AM): 152; SVM-​TIA 150; in time series 152;
anisotropic microstructure 178; UnRAP 153; VLNPS 148–​9
benefits 178; digital manufacturing audio processing 179
industrial paradigm 178 autism spectrum disorder (ASD) 80–​1
AI and ML process: aggressors 244; autonomous driving 179
amalgam of 244; reinforcement autoregressive integrated moving
learning scheme 246; security and average (ARIMA) model 163;
privacy 244–​5, 245; supervised bandwidth utilization 223; difference
learning scheme 246; unsupervised in bandwidth 225, 226; feature
learning scheme 246 extraction selection 223–​4, 225;
airport operations 168 forecasted values of bandwidth 232–​4,
algorithm process, ML 37 232–​4; model building 231; model
alleged robo-​scientists 66 evaluation 234, 234; time series data:
Amazon Echo device 3 maximum number of users 225, 227;
Amazon Mechanical Turk 81 seasonality 225, 226; seasonal patterns
anatomization method 197–​8 227–​30; TX speeds 225, 226
application-​specific integrated circuits autoregressive (AR) model 168,
(ASICs) 60, 68 258–​9
applied support vector machines 257–​8 autoregressive moving average (ARMA)
artificial intelligence (AI) 2, 36; model 256, 259–​60
description 20; EV users 56; large average absolute error (AAE) 188
government surveys 56; national average banks 49–​50
evidence 57; policy evaluation 57;
researchers 57; science and technology back-​propagation neural network
(S&T) research 57; supercharge battery (BPNN) 21
development 58, 62–​5; see also reverse bagging method 46
engineering Banco Bilbao Viscera Argentaria (BBVA)
artificial neural networks (ANNs) 20 50–​1
ASUS Xtion framework 75 battery electric vehicles (BEVs) 67
attack items: DMD and HIN 151; battery management system (BMS) 60,
dynamic time interval segmentation 68–​9
technique 153 battery testing 58–​9
attack profiles: classification 147, 147–​8; beamforming 246–​7
data tracking detection technique behavior prediction 278–​80

315
316

316 Index

bias: demographic blindness 43; 207; non-​membership correlation 202;


demographic parity 43; equal odds 44; privacy-​preserving data publishing
equal opportunity 44 models 200; quasi-​identifiers
binary decision tree (BDT) 150–​1 correlation 202; quasi-​identifier table
bioinformatics: advanced computational 203, 203–​5; sensitive attribute table
approaches 131; cellular mechanisms, 204, 204, 205–​6; see also information
analysis of 132; coronavirus SARS-​ loss
CoV-​2 132; description 131; proteins channel coding 247
131–​2; see also functional annotation of chemical hazard 37
proteins Cisco Annual Internet Report
biomedical technology 179 (2018–​2023) 240
BLAST algorithm 134 (c,k)-​anonymization 198
boosting method 46 classification technique 45
brain–​computer interface (BCI): classroom monitor system 81–​2
biomedical engineers and researchers cluster-​based classifiers 78
100; command generation 104; cognitive radio 247–​8
communication system 103; defined collaborative filtering (CF)-​based system
101; ECG 101–​2; EEG 101; MEG 102, 142; description 142, 143; IBCF 144–​5,
103; modalities and procedure 100; 145; UBCF 143–​4, 144
MRI 102; pattern classification 104; collaborative filtering recommender
pattern recognition 103–​4; signal system (CFRS) 142
acquisition 103 collective privacy preservation 196
breast cancer: classification 117; data command generation 104
description 122; data preprocessing communicative gestures 74
119; decision tree 121; feed-​forward computational algorithm 36
neural network 118; GONN 118; computer 108–​9
heat map of correlation 123; IoT and computer-​aided detection (CAD) 4
AI 118; K-​Nearest neighbors (KNN) computer vision (CV) 179
120; linear regression 120; logistic content-​based recommendation 142
regression 120–​1; ML, evaluation of contrastive learning: method 300,
122; random forest classification 122; 300–​1; ResNet-​18 model 299;
SVM 121; training of ML models 119 SimCLR framework 298, 299; visual
bucket-​ individual multi-​sensitive representations 298; without human
attribute bucket (QIAB-​IMSB) 198 supervision 298
business world: bias 43–​4; compliance controlling gestures 74
risk 42; comprehensive business conversational gestures 74, 77, 81
plans 41; dynamic model calibration convolutional architecture for fast
44; feature engineering 44; financial feature embedding (Caffe) 108
risk 42–​3; hyper parameters 44; convolutional neural network (CNN) 75,
interpretability 43; operational risk 42; 106, 179, 256, 261, 284
production readiness 44; reputational cooperative human–​machine diagnostic
risk 43; strategic risk 42 techniques 5
correct classification rate (CCR) 137
Cancer Dependency Map 2 credit risk 48
cancer models 2 credit risk management system (CRMS)
CBPP algorithm: anonymity 201, 202; 48
background knowledge correlation critical assessment of functional
200; batch id, bid 204–​5; evaluation annotation (CAFA) 136
317

Index 317

cyber-​physical systems (CPS) 178 discrete wavelet transform and the


cytotoxicity 277 support vector machine (DWTSVM)
149
Dangaard–​Oeschger (DO) events 166 DNNs 260–​1
data analysis MI: Attribute 117 as double M detector (DMD) technique 151
A_​V_​BLOK 291; Attribute 120 as 3DRCNN, hybrid framework 84
DRESSLER 291; Attribute 116 as dream, theories of: activation synthesis
FIBR_​JELUD 291; Attribute 113 as theory 95–​6; description 94, 95;
FIBR_​PREDS 290; Attribute 115 as information-​processing theory 95;
JELUD_​TAH 290–​1; Attribute 124 as physiological function theory 96;
LET_​IS 292; Attribute 118 as OTEK_​ wish-​fulfillment theory 95; see also
LANC 291; Attribute 123 as P_​IM_​ origins of dream
STEN 292; Attribute 114 as PREDS_​ driven EVs 69–​70
TAH 290; Attribute 119 as RAZRIV DW based on variance change (DWVC)
291; Attribute 122 as REC_​IM 292; 168
Attribute 121 as ZSN 291 dynamic model calibration 44
data anonymization 196 dynamic time interval segmentation
data preprocessing 119 technique 153
data publisher 196 dynamic time wrapping/​warping
data tracking detection technique 153 (DTW) 75, 78
DATimeS software 162 dynamic window (DW) based on label
DC–​DC converter 70 change (DWLC) 168
decision support systems 168
decision tree algorithm 118, 119, 121, electrical demand forecasting: data
124, 125 preprocessing 262; dataset 261–​2;
deep-​belief network (DBN) 105 multivariate time-​series forecasting
deep learning (DL) method 75; artificial 264, 265; univariate time-​series
neural networks (ANNs) 2; Caffe 108; forecasting 262, 264
CNN 105, 106; DBN 105; GAN 106, electric drive (E-​drive) 67
107; human nervous system 105; Keras electricity: extreme price volatility 256;
107–​8; MATLAB 109, 110; medical power price 256; sources 256
imaging 107; models 105; PyTorch electricity price forecasting 256
framework 108–​9; radiologists 4; electric vehicles (EVs): battery
recurrent neural network (RNN) 106; performance 56, 58; battery testing
RNN 105, 106 58–​9; production and deployment
deep neural network learning 179, 183–​4 62–​3; wider applications 59–​60; see also
degradation rate 277 artificial intelligence (AI)
demographic blindness 43 electric vehicles/​hybrid electric vehicles
demographic parity 43 (EV/​HEV) 70
dependent variable 104 electrocardiography (ECG) 101–​2
direct torque control (DTC) electroencephalography (EEG) 101
methodology 70 electronic health records (EHRs):
disaster management: Banco Bilbao literature’s unstructured data and
Viscera Argentaria (BBVA) 50–​1; metadata 6–​7; patient datasets 6
defined 50; machine learning (ML), elongation rate 277
role of 50 E-​Mobility: advantages of EVs 67;
discrete cosine transform (DCT) method auxiliary fuel hotspot, transportation
86 67; battery management systems
318

318 Index

(BMS) 68; conventional power sources biodegradable 275; bio-​toxic 275;


67; LiBs 60; sources of information 60 bone healing procedure 272; chemical
end-​user gesture analysis customization degradation process 274; corrosion
79 properties 275; hydroxyapatite
energy-​efficient network 247 (HA) composites 275; mechanical
energy stockpiling framework (ESS) properties 275–​6; non-​biodegradable
innovation 68 implants 272; orthopedic biomaterials
energy storage systems (ESS) 67 272; parameters 276; process of 271;
entropy 121 properties 273; surface degradation
equal odds 44 275; types of 272, 272
equal opportunity 44 fraud detection 48
evolutionary algorithm (EPUNN) 165 f-​slicing 198–​9
evolutionary ANNs (EANNs) 163–​4 fuel cell electric vehicles (FCEVs) 67
expectation maximization (EM) 150 functional annotation of proteins:
automatic 134–​7, 135, 136; NGS 134; in
feature engineering 44 vitro and in vivo experiments 133
federated learning 3, 9 fuzzy interference system (FIS) 21
field programmable gate arrays (FPGAs) fuzzy logic 21
60, 68
financial sector: conduct and market gaming 1
abuse 48–​9; credit risk 48; fraud Gaussian process regression (GPR) 162
detection 48 Gene Ontology (GO) Consortium 133
fine needle aspirate (FNA) 117 generative adversarial networks (GAN)
fixed window (FW) 168 250
flood risk mapping: adaptive neuro genetic algorithm (GA) 166
fuzzy inference system (ANFIS) 21; genetic algorithm–​support vector
artificial neural networks (ANNs) 20; machine (GA–​SVM) 258
description 19; LULC map creation 24, genetically optimized neural network
24; Markov model 25, 25; multilayer (GONN) 118
perception 20; remote sensing 24; state geographical information system (GIS)
selection 24; study area 21, 23, 23; technology 18
support vector machine (SVM) 21, gestures: autism spectrum disorder
22–​3; transition matrix calculations (ASD) 80–​1; capturing/​acquiring
25–​7, 26; wavelet neuro networks images 76; classroom monitor system
(WNN) 21 81–​2; end-​user gesture analysis
floods: emergency management customization 79; feature extraction
18; governments 19; human lives challenges 77; hand gestures 74;
devastation 18; image processing and human–​computer interaction 75;
predictive analysis 18 learning frameworks 83; learning
fog formation 167–​169, 169 individual styles 81; limitations and
8-​fold-​cross-​validation strategy 189 challenge 78; LSTM model 84; and
fracture healing: artificial replacements movement research 82; multi-​order
272, 273; bioactive 275; biodegradable feature analysis 79–​80; naturalistic
implants 272; biodegradation settings 80; sensory receptor devices
progression 274; bioinert 275; 75; tracking, segmentation and
biomaterials, classification of 274; identification challenge 76–​7; see also
biomechanical behavior, orthopedic capturing/​acquiring images
implants 276; bioresorbable/​ 5G network slicing 247
319

Index 319

Google Home device 3 information gain 121


government sectors 196 information loss, query accuracy 207–​8,
gradient-​adversarial network (GAN) 208
105, 106, 107 information-​processing theory 95
graphic processing units (GPUs) 60, 68 INFORMS dataset 198, 199
Greenland Ice Sheet Project Two (GISP2) intelligent transportation systems (ITS)
166 68
5G technology: approaches 248; International Business Machines Corp.
complexity 250; customer satisfaction (IBM) 65
249; “game-​evolving” innovation 240; internet of things (IoT)-​based methods
hyper-​connectivity 243; improved 70, 178
versatile broadband 241; and IoT 241; internet service providers (ISPs):
key enablers 241, 242; low latency clients/​servers 214; data collection
with ultra-​reliability 241; M2M method 215–​16, 215–​16; data set
communication 240, 241; performance information 216–​17; data visualization
of network 249; planning of network with tableau 218, 219–​20; network
249; product life cycle 249; RAT 243; topology 217, 217–​18; network traffic
requirements 243; revenue generation classification 214; preprocessing 218;
249; security 250; security and privacy see also traffic classification
251; trade-​off in speed 251; versatile interpretability: linear monotonic
network 240; see also AI and ML models 43; nonlinear monotonic
process models 43; nonlinear no monotonic
model 43
hand gestures 74 item-​based collaborative filtering (IBCF)
Hartanto’s technique 83 algorithm 144–​5, 145
hazard identification 39, 39–​40 iterative error feedback (IEF) 86
health care: limitations 9–​10; medical
diagnostics 3–​5, 4; monitoring 7, 7–​9; Jupyter notebook environment 123
precision health 5–​7
health sectors 196 k-​anonymity 198
heterogeneous information network Keras, open-​source framework 107–​8
(HIN) 151 kernel discriminant learning for ordinal
hidden Markov models (HMMs) 75, 152 regression (KDLOR) 168
hierarchical clustering 152 (k, km)-​anonymous method 197
high-​definition satellite images 19 k-​means classification approach 78
high-​fidelity imaging scans 3 K-​Nearest neighbors (KNN) algorithm
Hilbert–​Haung transform and support 118, 120, 123–​4, 124, 144
vector machine (HHT-​SVM) 148
human visual cortex 98, 99 land cover 18
hybrid recommendation approaches 142 land use 18
hydroxyapatite (HA) composites 275 land use and land cover (LULC) pattern
hyper parameters 44 18
l-​diversity 198
ImageNet 9 leap motion controller 75
independent variable 104 light detection and ranging (LIDAR) 62
individual privacy 196 likelihood ratios (LRs) 288
Industry 4.0. Smart manufacturing 178 linear classifier code 308
inertial measurement units (IMUs) 62 linear monotonic models 43
320

320 Index

linear regression (LR) 21, 120 myocardial infarction (MI): datasets


lithium-​ion batteries (LiBs) 60, 65, 68–​9 289, 290; early stage-​diagnosis 284;
logistic regression algorithm 118, 120–​1, outcomes prediction 284; research 289;
126 silent killer 284; symptoms 288; see also
long short-​term memory (LSTM) 163–​4, neural decision forests (NDFs)
256, 260
loss function 187 Naïve Bayes (NB) algorithm 150, 222
low-​thickness equality check (LDPC) National Data Buoy Centre (NDBC) 164
codes 247 natural language processing (NLP) 179
Lsl-​diversity model 198 NDWI Map of year 2015 29
LSMVHandPoses dataset 85 network traffic classification method
LULC map creation 24, 24 234–​5
neural decision forests (NDFs):
machine-​to-​machine (M2M) architecture 285; CNNs and random
communication 240 forest 284; cohorts 286; deep 285;
magnetic resonance imaging (MRI) 102 performance 294; types of weights
magnetoencephalography (MEG) 102, 285–​6; XGBoost 286
103 neural language processing (NLP) 108–​9
manipulative gestures 74 neural network (NN) 21, 179
man-​made intelligence 65 neurofuzzy 21
manufacturing 177–​8 new energy vehicles (NEVs) 67
Markov model 25, 25, 27, 30, 31 next-​generation sequencing (NGS) 134
Massachusetts Institute of Technology nonlinear activation functions 20
(MIT) 57 nonlinear monotonic models 43
massive MIMO (M-​MIMO) 246–​7 nonlinear no monotonic model 43
MATLAB software 28, 31, 109, 110 non-​sensitive attributes 196
Matplotlib (data visualization) 117–​18 nontechnical members 45
mean absolute error (MAE) 256–​7 normalized difference water index
mechanical hazard 37 (NDWI) 24
memory vs. dream: defined 97; encoding North Greenland Ice Core Project
98; long-​term memory formation 98; (NGRIP) 166
temporary 97–​8 novel shilling attack detection 150–​1
Microsoft Kinect framework 75 NREM sleep 96, 97
model training: neural decision forest NT-​Xent Loss function code 308, 312
model 293; neural decision tree model numerical weather prediction (NWP)
293; parameters 292 models 167
modulus of elasticity 277 NumPy (mathematical operations on the
molecular and phenotypic datasets 1 data set) 117–​18
molecular assays 4
moving average (MA) model 259 oceanography buoys 164
multi-​label problem 135 ontology 133
multilayer perceptrons (MLPs) 20, 164 origins of dream: cortical regions
multi-​objective evolutionary algorithm hippocampus and amygdala 96;
(MOEA) 168 hippocampus and amygdala 99,
multi-​order feature analysis 79–​80 99–​100, 100; human visual cortex 98,
multivariate forecasting models: CNN 99; vs. memory 97–​8, 98; vs. sleep 96–​7,
261; CNN-​LSTM architecture 261; 97, 97
CNN-​LSTM-​DNN 261; DNNs 260–​1 overfitting 187–​8
multivariate models 266, 268 overlapped slicing 197
321

Index 321

palaeoclimatology 165 rapid eye movement (REM) sleep stage


Pandas (data manipulation and analysis) 96, 97
117–​18 rating deviation from mean agreement
pattern classification 104 (RDMA) technique 150
pattern recognition 103–​4 recall 122
percentage of correct key points (PCKh) recommender systems (RSs): CF (see
86 collaborative filtering (CF)); “deceitful
pharmaceutical agencies 196 producers” 143; PPCF systems 143;
physiological function theory 96 see also profile injection attack method;
PIC microcontroller 87 shilling profiles (SPs)
(p,k)-​anglicization 197 rectified linear unit (ReLU) 186
plug-​in hybrid vehicle (PHEVs) 67 recurrent neural network (RNN) 75, 106,
polynomial function 121, 126, 128 164, 256
powder bed fusion (PBF) 180–​1, 181 regression technique 45
precision health 122; digital frameworks remaining useful life (RUL) 60, 68
6; molecular biomarkers 5–​6; remote sensing 18, 19, 24
multiscale predictive frameworks rendered hand pose dataset (RHD) 85
6; patient-​centered laboratory re-​scale AdaBoost (RAdaBoost) 148
frameworks 6 ResNet network 184, 308, 308–​9, 312
predictive intelligent battery reverse engineering: AI self-​driving
management system (PIBMS) 68 vehicle 62; CAD framework 61–​2;
predictive modeling techniques 19 manufacturers 61; vehicle creators 61
principal component analysis (PCA) 149 risk analysis: assessment 39; business
privacy-​preserved data publishing 197 world (see business world);
privacy-​preserving CF (PPCF) systems communication 39; components of 39;
143 hazard identification 39–​40;
private health information (PHI) 9 management 39; qualitative
production readiness 44 and quantitative methods 38;
profile injection attack method 145, 146 semiquantitative methods 38;
proportional odds model (POM) 168 types of hazard 37; unwanted/​
PyTorch framework 108–​9 harmful event 37
risk assessment 39
QI-​MHSA generalization algorithm risk communication 39
198 risk estimation 40
qualitative methods 38 risk management 39–​41, 47; accuracy
quantitative methods 38 of machine learning techniques 47;
quantum geographical information availability of data 47; availability of
system (QGIS) 24, 27 skilled staff 47; case studies: average
query accuracy 207–​8, 208 banks 49–​50; disaster management
50, 50–​1; transparency and ethics 47;
radar sensors 62 see also financial sector
radial base functions (RBFs) 168 risk prediction models (RPMs) 286–​7
radial basis function 121 runway visual range (RVR) 168
radio access network (RAN) 247
radio-​access technology (RAT) 243 science and technology (S&T) research
rainfall data: of year 2011 27, 28; 57
of year 2015 30 Seaborn (making heatmaps) 117–​18
random forest 128, 284 selective laser sintering (SLS):
random tree walks (RTW) 86 contraction ratio 185; dataset 182;
322

322 Index

deep neural network parameters state-​based transition model 24


185–​7; description 178–​9; K-​fold-​ state of charge (SoC) 60, 68
cross-​validation 187; models 182; state of health (SoH) 60, 68
overfitting 187–​8; parameters 184; statistical detection algorithm 152
parameters with levels 185, 186; PBF stereo hand pose tracking benchmark
process 180–​1, 181; semisupervised (STB) 85
learning 183; shrinkage ratio 185; supercharge battery development 58,
supervised learning 183; training 182; 62–​5
unsupervised learning 183 superconducting quantum interference
self-​driven cars 1 device (SQID) 102
self-​supervised learning method 300, 300 supervised learning 45, 46, 183
semiquantitative methods 38 support vector machine (SVM) 21, 22–​3,
semisupervised learning 183 78, 118, 119, 121, 126–​8
semi-​supervised learning based support vector machine and target item
shilling attack detection (SEMI-​SAD) analysis (SVM-​TIA) 150
technique 150 support vector ordinal regression
sensitive attributes 196 considering explicit constraints
shilling profiles (SPs): advantages (SVOREX) 168
and disadvantages 154–​7; detection support vector ordinal regression
146; elements of 145–​6; frequency considering implicit constraints
reduction 145; general 146; rating 146; (SVORIM) 168
rating and time interval parameter
151; rating parameter 147; size of 146; t-​closeness 198
time interval 146; see also attack items; technical hazard 37
attack profiles technical members 45
short-​term average change activity tensile strength 277
(SACA) 153 terminal aerodrome forecasts (TAFs) 170
short-​term variance change activity three-​D (3D) hand pose estimation
(SVCA) 153 79–​80
shrinkage ratio 185 time series data mining (TSDM):
sigmoid function 121, 127, 128 classification 161; clustering
sigmoid units (SUs) 168 algorithms 163; DATimeS 162; defined
signal acquisition 103 162, 163; fog formation 167–​169,
sign language 74; see also gestures 169; missing values imputation 162;
SimCLR framework 298; augmented piecewise ; inear approximations 163;
versions 301; data augmentation segmentation procedure 162; tipping
302–​3, 303, 304, 309, 309–​10; dataset points (TPs) 165–​7; wave height time
302; neural network encoder 303, 305, series 164–​5
305; nonlinear projection head 306, time series prediction 235
306–​7; normalized temperature-​scaled tipping points (TPs) 165–​7
cross-​entropy loss function 307–​8; traffic classification: classifiers 222, 222;
ResNet-​18 training 308, 308–​9 cleaned data 222; network model 220,
Sklearn package (normalization of the 221
data set) 117–​18 training losses 311–​12, 311–​12
sleep gesture detection 81–​2 training process 278
sleep vs. dream 96–​7, 97, 97 transition probability matrix 25
sonic sensors 62 transmission service operator (TSO)
speech-​based home aids 8 261–​2
323

Index 323

t-​SNE visualizations: last layer vectors wave height time series: oceanography
310; second last layer vectors 311 buoys 164; two-​staged procedure
165
UniProtKB/​Swiss-​Prot consortium 134 wavelet neuro networks (WNN) 21
univariate models 264–​6, 266 wavelets transform (WT) 21
unsupervised learning 45, 46, 183 wireless AI: channel coding 247;
unsupervised retrieval of attack profiles cognitive radio 247–​8; energy-​efficient
(UnRAP) detection technique 153 network 247; 5G network slicing 247;
user-​based CA algorithm 142 M-​MIMO and beamforming 246–​7;
user-​based collaborative filtering modulation regulation 248
(UBCF) 143–​4, 144 wish-​fulfillment theory 95

variable-​length partitions with neighbor yield strength 277


selection (VLPNS) algorithm 148–​9 YouTube dataset 198, 199
324

You might also like