Data‑Centric Artificial ­Intelligence

for Multidisciplinary Applications

This book explores the need for a data‑centric AI approach and its application in
the multidisciplinary domain, compared to a model‑centric approach. It examines
the methodologies for data‑centric approaches, the use of data‑centric approaches
in different domains, the need for edge AI and how it differs from cloud‑based AI.
It discusses the new category of AI technology, “data‑centric AI” (DCAI), which
focuses on comprehending, utilizing, and reaching conclusions from data. By add‑
ing machine learning and big data analytics tools, data‑centric AI modifies this by
enabling it to learn from data rather than depending on algorithms. It can there‑
fore make wiser choices and deliver more precise outcomes. Additionally, it has the
potential to be significantly more scalable than conventional AI methods.

• Includes a collection of case studies with experimentation results to adhere

to the practical approaches
• Examines challenges in dataset generation, synthetic datasets, analysis, and
prediction algorithms in stochastic ways
• Discusses methodologies to achieve accurate results by improving the
­quality of data
• Comprises cases in healthcare and agriculture with implementation and
impact of quality data in building AI applications
 Data‑Centric Artificial
Intelligence for
Multidisciplinary
Applications

Edited by
Parikshit N. Mahalle, Namrata N. Wasatkar, and
Gitanjali R. Shinde
Designed cover image: ShutterStock

First edition published 2024

by CRC Press
2385 NW Executive Center Drive, Suite 320, Boca Raton FL 33431

and by CRC Press

4 Park Square, Milton Park, Abingdon, Oxon, OX14 4RN

CRC Press is an imprint of Taylor & Francis Group, LLC

© 2024 selection and editorial matter, Parikshit N. Mahalle, Namrata N. Wasatkar and Gitanjali R.
Shinde; individual chapters, the contributors

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.copyright.
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 mpkbookspermissions@
tandf.co.uk

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: 9781032610061 (hbk)

ISBN: 9781032610078 (pbk)
ISBN: 9781003461500 (ebk)

DOI: 10.1201/9781003461500

Typeset in Times
by codeMantra
Contents
Editors..................................................................................................................... viii
List of Contributors.....................................................................................................x

Section I  R
 ecent Developments in
Data‑Centric AI
Chapter 1 Advancements in Data‑Centric AI Foundations, Ethics,
and Emerging Technology....................................................................3
Sujal Dilip Patil, Rupali Atul Mahajan, and Nitin Sakhare

Chapter 2 Emerging Development and Challenges in Data‑Centric AI.............. 27

Chaitali Shewale

Chapter 3 Unleashing the Power of Industry 4.0: A Harmonious

Blend of Data‑Centric and Model‑Centric AI.....................................40
Manivannan Karunakaran, Batri Krishnan, D. Shanthi,
J. Benadict Raja, and B. Sakthi Karthi Durai

Chapter 4 Data‑Centric AI Approaches for Machine Translation....................... 55

Chandrakant D. Kokane, Pranav Khandagale, Mehul Ligade,
Shreeyash Garde, and Vilas Deotare

Section II  D
 ata‑Centric AI in
Healthcare and Agriculture

Chapter 5 Case Study Medical Images Analysis and Classification

with Data-Centric Approach............................................................... 79
Namrata N. Wasatkar and Pranali G. Chavhan

Chapter 6 Comparative Analysis of Machine Learning Classification

Techniques for Kidney Disease Prediction......................................... 88
Jayashri Bagade, Nilesh P. Sable, and Komal M. Birare

v
vi Contents

Chapter 7 Fusion of Multi‑Modal Lumber Spine Scans Using

Convolutional Neural Networks..........................................................99
Bhakti Palkar

Chapter 8 Medical Image Analysis and Classification for Varicose Veins....... 114
Jyoti Yogesh Deshmukh, Vijay U. Rathod, Yogesh Kisan Mali,
and Rachna Sable

Chapter 9 Brain Tumor Detection Using CNN.................................................. 132

Paras Bhat, Sarthak Turki, Vedyant Bhat, Gitanjali R. Shinde,
Parikshit N. Mahalle, Nilesh P. Sable, Riddhi Mirajkar, and
Pranali Kshirsagar

Chapter 10 Explainable Artificial Intelligence in the Healthcare:

An Era of Commercialization for AI Solutions................................ 142
Prasad Raghunath Mutkule, Nilesh Popat Sable,
Parikshit N. Mahalle, and Gitanjali R. Shinde

Chapter 11 Role of Data‑Centric Artificial Intelligence in Agriculture.............. 159

Rajkumar Patil, Nilesh Popat Sable, Parikshit N. Mahalle,
Gitanjali R. Shinde, Prashant Dhotre,
and Pankaj Chandre

Chapter 12 Detection and Classification of Mango Fruit‑Based on Feature

Extraction Applying Optimized Hybrid LA‑FF Algorithms............ 177
Mukesh Kumar Tripathi, M. Neelakantappa,
Parikshit N. Mahalle, Shylesha V. Channapattana,
Ganesh Deshmukh, and Ghongade Prashant

Section III  B
 uilding AI with Quality Data
for Multidisciplinary Domains

Chapter 13 Guiding Your Way: Solving Student Admission Woes..................... 189

Snehal Rathi, Shekhar Chaugule, Manisha Mali,
Gitanjali R. Shinde, and Swati Patil
Contents vii

Chapter 14 Melodic Pattern Recognition for Ornamentation Features

in Music Computing..........................................................................202
Makarand Ramesh Velankar, Sneha Kiran Thombre,
and Harshad Suryakant Wadkar

Chapter 15 Content Analysis Framework for Skill Assessment.......................... 214

Abhishek Kabade, Harshad Jagadale, Anurag Bharde,
and S. P. Sonavane

Chapter 16 Machine‑Learning Techniques for Effective Text Mining............... 234

Shivam Singh, Chandrakant D. Kokane, Vilas Deotare,
and Tushar Waykole

Chapter 17 Emails Classification and Anomaly Detection using Natural

Language Processing......................................................................... 271
Tanvi Mehta, Renu Kachhoria, Swati Jaiswal, Sunil Kale,
Rajeswari Kannan, and Rupali Atul Mahajan

Index....................................................................................................................... 293
Editors
Dr. Parikshit N. Mahalle i s a senior member of the IEEE and
is Professor and Head of Department of Artificial Intelligence
and Data Science at Vishwakarma Institute of Information
Technology, Pune, India. He completed his Ph.D. from Aalborg
University, Denmark and continued as Post Doc Researcher at
CMI, Copenhagen, Denmark. He has 23+ years of teaching
and research experience. He is a member of the Board of
Studies in Computer Engineering, Ex‑Chairman Information
Technology, SPPU and various Universities and autonomous
colleges across India. He has 9 patents, 200+ research publications (Google Scholar
citations‑2250 plus, H index‑22 and Scopus Citations are 1190 plus with H index ‑16),
and authored/edited 42+ books with Springer, CRC Press, Cambridge University
Press, etc. He is editor‑in‑chief for IGI Global – International Journal of Rough Sets
and Data Analysis, Associate Editor for IGI Global – International Journal of
Synthetic Emotions, Inter‑science International Journal of Grid and Utility
Computing, member of Editorial Review Board for IGI Global – International
Journal of Ambient Computing and Intelligence. His research interests are machine
learning, data science, algorithms, internet of things, identity management and secu‑
rity. He is a recognized Ph.D. guide of SSPU, Pune, guiding seven Ph.D. students in
the area of IoT and machine learning. Recently, five students have successfully
defended their Ph.D. He is also the recipient of the “Best Faculty Award” by Sinhgad
Institutes and Cognizant Technologies Solutions. He has delivered 200 plus lectures
at national and international levels. He is also the recipient of the best faculty award
by Cognizant Technology Solutions.

Dr. Gitanjali R. Shinde is Head and Associate Professor in

the Department of Computer Science & Engineering (AI
&ML), Vishwakarma Institute of Information Technology,
Pune, India. She completed her Ph.D. in Wireless
Communication from CMI, Aalborg University, Copenhagen,
Denmark on the Research Problem Statement “Cluster
Framework for Internet of People, Things and Services” – Ph.D.
awarded on 8 May 2018. She earned her M.E. in Computer
Engineering from the University of Pune, Pune, in 2012 and
B.E. in Computer Engineering degree from the University of Pune, Pune, in 2006.
She received research funding for the project “Lightweight group authentication for
IoT” by SPPU, Pune. She presented a research article in the World Wireless Research
Forum (WWRF) meeting, Beijing, China. She received the best paper award at an
international conference. She is also reviewer of various international journals
Springer, IGI Global, IEEE Transaction and various conferences. She has published
50+ papers in National and International conferences and journals (Google Scholar

viii
Editors ix

citations‑700 plus, H index‑11). She is author of 10+ books with publishers Springer
and CRC Press Taylor & Francis Group, and she is also editor of several books. Her
book Data Analytics for Pandemics A COVID 19 Case Study was awarded outstand‑
ing Book of the year 2020.

Dr. Namrata N. Wasatkar is an Assistant Professor in the

Department of Computer Engineering, Vishwakarma Institute
of Information Technology, Pune, India. She did her Ph.D. in
Computer Engineering from Savitribai Phule Pune University,
Pune, India on the research problem statement “Rule based
Machine translation of simple Marathi sentences to English
sentences” – Ph.D. awarded on 17 November 2022. She earned
an M.E. in Computer Engineering from the University of Pune,
Pune, in 2014 and B.E. in Computer Engineering from the
University of Pune, Pune, in 2012. She has received research
funding for the project “SPPU online chatbot” by SPPU, Pune. She is also reviewer
for various journals and conferences. She has published 15+ papers in National and
International conferences and journals. She has authored a book titled Data Centric
Artificial Intelligence: A Beginner’s Guide.
Contributors
Jayashri Bagade Pranali G. Chavhan
Department of Information Technology Vishwakarma Institute of Information
Vishwakarma Institute of Information Technology
Technology Pune, India
Pune, India
Vilas Deotare
Anurag Bharde PCET’s Nutan Maharashtra Institute of
Department of Information Technology Engineering and Technology
Walchand College of Engineering Talegaon Dabhade, India
Sangli, India
Ganesh Deshmukh
Paras Bhat Department of Computer Science and
Vishwakarma Institute of Information Engineering
Technology Vardhaman College of Engineering
Pune, India Hyderabad, India

Vedyant Bhat Jyoti Yogesh Deshmukh

Vishwakarma Institute of Information Marathwada Mitramandal’s Institute of
Technology Technology
Pune, India Pune, India

Komal M. Birare Prashant Dhotre

Vishwakarma Institute of Information Department of Information Technology
Technology MIT Art, Design and Technology
Pune, India University
Pune, India
Pankaj Chandre
Department of Information Technology B. Sakthi Karthi Durai
MIT Art, Design and Technology School of Computing
University Kalasalingam Academy of Research
Pune, India and Education
Srivilliputhur, India
Shylesha V. Channapattana
Department of Mechanical Engineering, Shreeyash Garde
KLS Vishwanathrao Deshpande PCET’s Nutan Maharashtra Institute of
Institute of Technology Engineering and Technology
Haliyal, India Talegaon Dabhade, India

Shekhar Chaugule Harshad Jagadale

Vishwakarma Institute of Information Department of Information Technology
Technology Walchand College of Engineering
Pune, India Sangli, India

x
Contributors xi

Swati Jaiswal Batri Krishnan

Department of Computer Engineering School of Computer Science and
Pimpri Chinchwad College of Engineering
Engineering JAIN (Deemed to be University)
Pune, India Ramnagara, India

Abhishek Kabade Pranali Kshirsagar

Department of Information Vishwakarma Institute of Information
Technology Technology
Walchand College of Engineering Pune, India
Sangli, India
Mehul Ligade
Renu Kachhoria PCET’s Nutan Maharashtra Institute of
Department of AI and DS Engineering and Technology
Vishwakarma Institute of Information Talegaon Dabhade, India
Technology
Pune, India Rupali Atul Mahajan
Vishwakarma Institute of Information
Sunil Kale Technology
Department of AI and DS Pune, India
Vishwakarma Institute of Information
Technology Parikshit N. Mahalle
Pune, India Vishwakarma Institute of Information
Technology
Rajeswari Kannan Pune, India
Department of Computer Engineering
Pimpri Chinchwad College of Manisha Mali
Engineering Vishwakarma Institute of Information
Pune, India Technology
Pune, India
Manivannan Karunakaran
School of Computer Science and Yogesh Kisan Mali
Engineering G H Raisoni College of Engineering
JAIN (Deemed to be University) and Management
Ramnagara, India Pune, India

Pranav Khandagale Tanvi Mehta

PCET’s Nutan Maharashtra Institute of Department of Computer Engineering
Engineering and Technology Pimpri Chinchwad College of
Talegaon Dabhade, India Engineering
Pune, India
Chandrakant D. Kokane
PCET’s Nutan Maharashtra Institute of Riddhi Mirajkar
Engineering and Technology Vishwakarma Institute of Information
Talegaon Dabhade, India Technology
Pune, India
xii Contributors

Prasad Raghunath Mutkule Snehal Rathi

Sanjivani College of Engineering Vishwakarma Institute of Information
Kopargaon, India Technology
Pune, India
M. Neelakantappa
Department of Information Technology Vijay U. Rathod
Vasavi College of Engineering G H Raisoni College of Engineering
Hyderabad, India and Management
Pune, India
Bhakti Palkar
K.J. Somaiya College of Engineering Nilesh Popat Sable
Pune, India Department of Computer Science &
Engineering (Artificial Intelligence)
Rajkumar Patil Vishwakarma Institute of Information
Department of Information Technology
Technology Pune, India
MIT Art, Design and Technology
University Rachna Sable
Pune, India G H Raisoni College of Engineering
and Management
Sujal Dilip Patil Pune, India
Vishwakarma Institute of Information
Technology Nitin Sakhare
Pune, India Vishwakarma Institute of Information
Technology,
Swati Patil Pune, India
Vishwakarma Institute of Information
Technology D. Shanthi
Pune, India Department of Computer Science and
Engineering
Ghongade Prashant PSNA College of Engineering and
Department of Electrical and Technology
Electronics Engineering Dindigul, India
Guru Nanak Dev Engineering College
Bidar, India Chaitali Shewale
Vishwakarma Institute of Information
J. Benadict Raja Technology
Department of Computer Science and Pune, India
Engineering
PSNA College of Engineering and Gitanjali R. Shinde
Technology Vishwakarma Institute of Information
Dindigul, India Technology
Pune, India
Contributors xiii

Shivam Singh Makarand Ramesh Velankar

PCET’s Nutan Maharashtra Institute of Information Technology Department
Engineering and Technology MKSSS’S Cummins College of
Talegaon Dabhade, India Engineering for Women
Pune, India
S. P. Sonavane
Department of Information Technology Harshad Suryakant Wadkar
Walchand College of Engineering Information Technology Department
Sangli, India MKSSS’S Cummins College of
Engineering for Women
Sneha Kiran Thombre Pune, India
Information Technology Department
MKSSS’S Cummins College of Namrata N. Wasatkar
Engineering for Women Vishwakarma Institute of Information
Pune, India Technology
Pune, India
Mukesh Kumar Tripathi
Department of Computer Science and Tushar Waykole
Engineering PCET’s Nutan Maharashtra Institute of
Vardhaman College of Engineering Engineering and Technology
Hyderabad, India Talegaon Dabhade, India

Sarthak Turki
Vishwakarma Institute of Information
Technology
Pune, India
Section I
Recent Developments
in Data‑Centric AI
 1 Advancements
in Data‑Centric AI
Foundations, Ethics, and
Emerging Technology
Sujal Dilip Patil, Rupali Atul Mahajan,
and Nitin Sakhare

1.1 INTRODUCTION
Data‑centric artificial intelligence (AI) denotes an approach within AI and
machine learning (ML) that places significant emphasis on the pivotal role of
meticulously curated, high‑quality data in the development and implementation of
AI models and systems [1]. Under this paradigm, data assumes the bedrock upon
which AI algorithms are constructed and honed, and its effective handling, pre‑
processing, and analysis stand as pivotal factors for achieving precise and depend‑
able AI outcomes [2]. The essence of data‑centric AI springs from the recognition
that the performance of AI models is intricately linked to the calibre and quantity
of data employed for training, validation, and testing. This methodology under‑
scores the understanding that even the most advanced AI algorithms might grap‑
ple to yield meaningful outcomes if the input data is incomplete, biased, laden
with noise, or inadequately structured [3]. The overall chapter consists of key
aspects of data‑centric AI described in Section 1.1.1, Applications of Data‑Centric
AI, Various techniques – Machine Learning, Deep Learning used in AI, etc. are
explained in Section 1.2. The various technologies that are part of data‑centric
AI are mentioned in Sections 1.3, 1.4, 1.5, and 1.6. The ethical implications of
AI technologies are described in Section 1.7. This chapter also covers various AI
Governance and Regulation Strategies for Responsible AI Implementation and
Oversight.
Key aspects of Data‑Centric AI include:

• Data Collection and Preparation: Careful and meticulous collection

of relevant and representative data is essential. Data needs to be cleaned,
pre‑processed, and transformed to ensure its suitability for training and
testing AI models. This process may involve dealing with missing values,
removing outliers, and addressing data imbalance.

DOI: 10.1201/9781003461500-2 3
4 Data-Centric Artificial Intelligence for Multidisciplinary Applications

• Data Quality: Ensuring data quality involves maintaining accuracy, con‑

sistency, and reliability throughout the data lifecycle. Quality control mea‑
sures should be in place to detect and correct errors that could impact AI
model performance.
• Data Bias and Fairness: Data‑centric AI acknowledges the potential for
bias in training data and strives to identify and mitigate biases that could
lead to unfair or discriminatory AI outcomes. Addressing bias requires
diverse and representative data as well as techniques for bias detection and
reduction.
• Data Augmentation: To enhance model generalization and robustness,
data augmentation techniques can be employed. These involve creating new
training examples by applying various transformations to existing data,
which helps models perform better on unseen data.
• Active Learning: Active learning strategies involve selecting the most
informative samples for annotation, reducing the amount of labelled data
needed to achieve a desired level of model performance [4]. This approach
optimizes the learning process and minimizes the labelling cost.
• Data Privacy and Security: Protecting sensitive and private data is para‑
mount. Data‑centric AI incorporates privacy‑preserving techniques to pre‑
vent unauthorized access, sharing, or misuse of sensitive information.
• Continuous Monitoring and Maintenance: The data‑centric approach
extends beyond model deployment. Continuous monitoring of model per‑
formance in real‑world scenarios helps identify potential issues or drift and
allows for timely model updates and retraining.
• Data Labelling and Annotation: Labelled data is often required for super‑
vised learning, where inputs are associated with corresponding desired
outputs. Data‑centric AI emphasizes the need for accurate and consistent
labelling, as errors in labelling can propagate through the training process
and affect model performance.
• Domain Expertise: Effective data‑centric AI requires collaboration
between AI practitioners and domain experts [5]. Domain experts provide
valuable insights into the nuances of the data and help guide the preprocess‑
ing, feature engineering, and model selection processes.
• Feature Engineering: Feature engineering involves selecting, transform‑
ing, or creating relevant features from the raw data to improve model perfor‑
mance. This process requires an understanding of the data’s characteristics
and the problem at hand.
• Data Versioning and Lineage: Maintaining a record of data versioning
and lineage is crucial for transparency and reproducibility. It enables track‑
ing changes to the data over time, which is essential for debugging and
ensuring consistent model behaviour [6].
• Data Governance: Establishing clear data governance policies ensures that
data is managed in a compliant and ethical manner. This includes defin‑
ing roles and responsibilities, access controls, and policies for data sharing
and usage.
Advancements in Data-Centric AI 5

• Data Pipelines: Building efficient data pipelines for data preprocessing,

augmentation, and model training streamlines the data‑centric AI work‑
flow [7]. Well‑organized pipelines contribute to consistent and reproducible
results.
• Scalability and Efficiency: As datasets grow in size and complexity,
data‑centric AI also focuses on scalable solutions that can handle large vol‑
umes of data efficiently. This might involve distributed computing, parallel
processing, and optimized storage [8].
• Feedback Loops: Integrating feedback loops into the AI process allows
models to continuously learn and adapt from real‑world usage. This iterative
approach helps models stay relevant and effective over time.
• Data Visualization and Interpretability: Visualizing data and model out‑
comes aids in understanding patterns and insights. Additionally, interpret‑
ability techniques provide insights into how models arrive at their decisions,
enhancing trust and transparency.
• Transfer Learning and Pretrained Models: Leveraging pretrained
models and transfer learning techniques can accelerate the training pro‑
cess and improve model performance, particularly when labelled data is
limited.
• Data‑Centric AI Ethics: Ethical considerations are paramount in data‑­
centric AI. This involves ensuring that the data used respects user privacy,
consent, and confidentiality and that the AI systems deployed adhere to
ethical guidelines [9].
• Data Federations: In situations where data cannot be centralized due
to privacy or security concerns, data federations allow AI models to be
trained across decentralized datasets while maintaining data ownership
and privacy.
• Quantifying Data Value: Data‑centric AI acknowledges that not all data
is equally valuable. Some data might have a greater impact on model per‑
formance than others. Techniques to quantify data value help allocate
resources efficiently during data collection and labelling.

Data‑centric AI represents a paradigm shift from focusing solely on algorithmic

advancements to recognizing the pivotal role of data in AI success. By addressing the
challenges and opportunities within the data lifecycle, this approach strives to create
AI models that are accurate, fair, transparent, and adaptable to real‑world scenarios.
These real‑world scenarios and use cases are described in Section 1.2.

1.2 APPLICATIONS OF DATA‑CENTRIC AI IN VARIOUS

DOMAINS
Data‑centric AI has a wide range of applications across various domains, where the
emphasis on high‑quality data and effective data management significantly enhances
the performance and impact of AI models. Here are some examples of how data‑cen‑
tric AI is applied in different fields:
6 Data-Centric Artificial Intelligence for Multidisciplinary Applications

1.2.1 Healthcare
Data‑centric AI improves the accuracy of medical image analysis, aiding in d­ iagnosing
diseases like cancer, detecting anomalies, and predicting patient ­outcomes [10].
AI‑driven analysis of molecular data accelerates drug discovery by predicting drug
interactions, identifying potential drug candidates, and optimizing drug designs [11].
By analysing patient data, including genetic, clinical, and lifestyle factors, AI helps
tailor treatment plans and predicts patient responses to therapies [12].

1.2.2 Finance
Data‑centric AI enhances credit risk assessment, fraud detection, and anti‑money
laundering efforts by analysing transactional and behavioural data for patterns. AI
models analyse market data to make high‑frequency trading decisions, optimizing
trading strategies, and portfolio management. AI‑driven analytics of customer data
improve marketing strategies, customer segmentation, and churn prediction [13].

1.2.3 Retail and E‑Commerce

Data‑centric AI predicts product demand, optimizing inventory management and
supply chain operations. AI‑powered recommendation engines analyse customer
behaviour to provide personalized product recommendations, enhancing user
­experience and sales. AI models analyse pricing data and market trends to optimize
pricing strategies and increase competitiveness.

1.2.4 Manufacturing
Data‑centric AI monitors and analyses sensor data from manufacturing processes to
identify defects, reduce waste, and ensure product quality. AI models analyse equip‑
ment sensor data to predict maintenance needs, minimizing downtime and opti‑
mizing maintenance schedules. AI‑driven analysis of supply chain data improves
inventory management, demand forecasting, and logistics planning.

1.2.5 Energy and Utilities

Data‑centric AI optimizes energy distribution by analysing consumption patterns,
predicting demand, and identifying areas for efficiency improvements. AI models
analyse weather and energy production data to optimize the integration and output of
renewable energy sources like solar and wind [14].

1.2.6 Transportation
Data‑centric AI powers self‑driving cars by analysing sensor data to make real‑time
driving decisions and ensure passenger safety [15]. AI‑driven analysis of traffic data
improves traffic flow, reduces congestion, and enhances urban mobility.
Advancements in Data-Centric AI 7

1.2.7 Agriculture
Data‑centric AI analyses data from sensors, drones, and satellites to optimize crop
management, irrigation, and fertilization for increased yield and resource efficiency.
AI models analyse images of plants to detect diseases early, enabling targeted inter‑
ventions and reducing crop loss.

1.2.8 Natural Language Processing (NLP)

Data‑centric AI analyses text data from social media and reviews to gauge public
sentiment about products, services, or events. AI‑powered translation models analyse
large volumes of text to provide accurate and context‑aware translations between
languages. These examples highlight how data‑centric AI is revolutionizing various
industries by leveraging high‑quality data to develop more accurate, efficient, and
insightful AI applications. The approach is adaptable to a wide range of domains,
each benefiting from improved decision‑making, enhanced customer experiences,
and increased operational efficiency.
The various technologies that are part of data‑centric AI are discussed below:

1.3 MACHINE LEARNING

ML is a branch of AI and computer science which focuses on the use of data
and algorithms to imitate the way that humans learn, gradually improving its
accuracy.

1.3.1 Supervised Learning
Supervised learning stands as a type of ML wherein the algorithm gains insights
from labelled training data. In this context, the dataset is composed of pairs ­denoting
inputs and their associated outputs. The algorithm’s objective revolves around
­mastering the mapping between inputs and the anticipated outputs. The ultimate
aim is for the algorithm to extrapolate from the training dataset and generate precise
­predictions or classifications for fresh, previously unseen data.
In supervised learning, the algorithm is furnished with datasets compris‑
ing labelled pairs of inputs and their corresponding outputs. The primary goal
of supervised learning is for the algorithm to discern and internalize a map‑
ping function. This function should enable the algorithm to make precise pre‑
dictions or classifications for new, previously unseen instances. Supervised
learning encompasses diverse tasks, with classification and regression standing
out. Classification entails assigning labels to inputs, while regression involves
predicting continuous values. A range of algorithms are utilized in supervised
learning, including decision trees, support vector machines, neural networks, and
linear regression. These algorithms are tailored to address various types of data
and tasks.
8 Data-Centric Artificial Intelligence for Multidisciplinary Applications

1.3.2 Unsupervised Learning
Unsupervised learning entails training an ML model using unlabelled data, where
explicit output labels are absent [16]. The algorithm’s objective is to uncover pat‑
terns, structures, or relationships inherent within the data. This is frequently accom‑
plished by clustering similar instances together or reducing the data’s dimensionality.
Unlabelled data is provided, and the algorithm seeks to identify underlying patterns
or groupings. The goal is to explore the inherent structure of the data, uncover hidden
patterns, or reduce its complexity. Unsupervised learning encompasses tasks such
as clustering, where similar data points are grouped together, and dimensionality
reduction, which involves condensing the number of features while preserving vital
information [17]. A variety of algorithms are employed in unsupervised learning,
including k‑means clustering, hierarchical clustering, principal component analysis,
and t‑distributed Stochastic Neighbor Embedding. These algorithms cater to differ‑
ent aspects of data exploration and pattern recognition.

1.3.3 Reinforcement Learning
Reinforcement learning (RL) stands as an ML category in which an agent learns to
formulate decisions through interactions with an environment. This agent garners
feedback in the shape of rewards or penalties contingent on its actions, striving to
acquire a policy that maximizes the cumulative reward across a span of time. RL
frequently finds application in tasks involving sequential decision‑making.

• Training Data: The agent learns through trial and error by interacting with
an environment.
• Objective: The agent aims to learn a policy that maximizes the expected
cumulative reward over a sequence of actions.
• Examples: Game playing, robotic control, and autonomous driving are
typical applications of RL.
• Algorithms: Q‑learning, Deep Q‑Networks, and policy gradient methods
are common RL algorithms.

These three fundamental learning algorithms constitute the bedrock of ML, serving as
pivotal components for constructing a diverse array of AI applications. Depending on
the nature of the challenge and the accessibility of labelled data, each learning algo‑
rithm category possesses its own merits and constraints. The selection of an algorithm
hinges upon the particular task at hand and the distinct attributes of the data involved.

1.4 DEEP LEARNING FUNDAMENTALS: NEURAL

NETWORK ARCHITECTURE AND DESIGN
Deep learning constitutes a subset of ML, focusing on the creation and training
of neural networks to acquire knowledge and formulate predictions from data.
Neural networks, inspired by the human brain’s structure and operation, serve as
Advancements in Data-Centric AI 9

computational models. These networks are comprised of interconnected nodes

(­neurons) organized into layers, capable of grasping intricate patterns and represen‑
tations from data through a process termed training.

1. Architecture of Neural Networks:

a. Input Layer: The initial layer accepts raw data, with each neuron cor‑
responding to a data attribute or feature.
b. Hidden Layers: Found between input and output layers, these inter‑
mediary layers house neurons that process input data through weighted
connections and activation functions [18].
c. Output Layer: The final prediction or output of the network emerges
from this layer. Neuron count in the output layer depends on the task,
like classification (number of classes) or regression (single output).
2. Neuron (Node) Function in Neural Networks:
i. A neuron within a neural network executes these steps:
a. Weighted Sum: Neuron inputs are multiplied by weights, and the
cumulative weighted inputs are computed.
b. Activation Function: The weighted sum undergoes an activation
function, introducing non‑linearity. ReLU, Sigmoid, and Tanh are
common activation functions.
3. Design Considerations:
a. Activation Functions: Choice of activation functions influences learn‑
ing and complex relationship modelling. ReLU is prevalent in hidden
layers due to its ability to address vanishing gradient issues.
b. Loss Function: Evaluating discrepancies between predictions and
actual targets, this function guides optimization in training. Distinct
tasks (classification and regression) necessitate distinct loss functions
(e.g., cross‑entropy and mean squared error) [19].
c. Optimization Algorithm: Optimization algorithms (e.g., Gradient
Descent and Adam) adapt network weights based on loss gradients.
They aim to minimize loss for enhanced model performance.
d. Regularization: Techniques like dropout, L1/L2 regularization, and
batch normalization counter overfitting by controlling network com‑
plexity and enhancing generalization.
e. Architecture Decisions: Determining layer count, neurons per layer,
and overall structure mandates experimentation, considering factors
like dataset size, complexity, and available resources [20].
4. Training and Backpropagation:
i. Training a neural network encompasses these steps:
a. Forward Pass: Input data traverses the network for predictions.
Comparing predicted and actual outcomes using the loss function
ensues [21].
b. Backpropagation: Utilizing the chain rule, gradients of loss con‑
cerning weights are computed. Updating weights in the opposite
gradient direction minimizes loss.
10 Data-Centric Artificial Intelligence for Multidisciplinary Applications

c. Iterative Optimization: Forward pass and backpropagation repeat

over multiple iterations (epochs) with data batches. The goal is to
identify weight configurations minimizing loss.
d. Validation and Testing: Model evaluation transpires on validation
and test datasets, gauging generalization performance.

Deep learning and neural networks have propelled advancements in diverse domains,
such as computer vision, NLP, and RL. Effective deep learning model development,
enabling precise predictions and discovery of intricate data patterns, demands pru‑
dent architecture design and rigorous training [22].

1.5 CONVOLUTIONAL NEURAL NETWORKS IN IMAGE ANALYSIS

Convolutional Neural Networks (CNNs) represent a specialized variant of neural
network architecture meticulously crafted for the examination and interpretation
of grid‑like data, encompassing images, video frames, and various structured data
formats. Renowned for their prowess in computer vision pursuits, CNNs excel in
tasks like image classification, object detection, image segmentation, and beyond.
The blueprint of CNNs takes inspiration from the visual processing mechanisms
inherent within the human brain. In convolutional layers, a collection of adaptable
filters (termed kernels) is applied to the input image. These filters discern a multitude
of features, encompassing edges, textures, or intricate patterns [23]. The convolution
operation entails sliding the filter across the input image. Element‑wise multiplica‑
tions and subsequent summation yield feature maps. Common activation functions
implemented in CNNs comprise ReLU (Rectified Linear Unit). ReLU introduces
non‑linearity to the network, thereby facilitating the modelling of intricate relation‑
ships. Pooling layers execute the down‑sampling of feature map spatial dimensions.
This action curtails computational intricacy while constructing a tiered portrayal
of features. Max pooling and average pooling are typical pooling methodologies,
wherein the maximum or average value within a pooling window is preserved.
Fully connected layers are used in the later stages of a CNN to make predictions
or decisions based on the extracted features. These layers connect every neuron to
every neuron in the previous and subsequent layers, like traditional neural networks.
Images are usually pre‑processed before feeding them into CNNs. This includes
resizing images to a consistent size, normalizing pixel values (scaling to a certain
range), and data augmentation (applying random transformations to increase data‑
set diversity). A typical CNN architecture consists of alternating convolutional and
pooling layers. Multiple convolutional layers with increasing complexity can capture
different levels of features.
After the convolutional layers, the feature maps are flattened into a vector and
passed through fully connected layers to perform classification or other tasks.
The output layer depends on the specific task. For image classification, it often
has soft‑max activation for multi‑class classification, providing probabilities for
each class. Transfer learning is a powerful technique in CNNs where pre‑trained
models (such as Visual Geometry Group ResNet, and Inception) are used as start‑
ing points for a specific task. The lower layers of these networks capture general
Advancements in Data-Centric AI 11

features, and fine‑tuning can be performed on higher layers to adapt the model to the
specific dataset.

1.5.1 Applications of CNNs
CNNs have demonstrated remarkable performance in various image analysis tasks:

• Image Classification: Assigning a label to an input image from a set of

predefined classes.
• Object Detection: Locating and classifying objects within an image.
• Image Segmentation: Assigning a label to each pixel in an image, distin‑
guishing different object regions.
• Facial Recognition: Identifying individuals from facial features.
• Medical Image Analysis: Diagnosing diseases from medical images
(X‑rays, MRIs, etc.).
• Autonomous Vehicles: Detecting pedestrians, vehicles, and obstacles in
real time.
• CNNs have revolutionized computer vision and image analysis, enabling
machines to understand and interpret visual information with increasing
accuracy and efficiency.

1.6 RECURRENT NEURAL NETWORKS FOR SEQUENTIAL DATA

Recurrent Neural Networks (RNNs) denote a subset of neural networks tailored to
manage sequential data, wherein the arrangement of elements holds significance.
Their proficiency notably shines in domains like NLP, speech recognition, time series
analysis, and any context‑entailing sequences. RNNs boast a capacity for internal
memory retention, thus rendering them adept at apprehending temporal dependen‑
cies and patterns. RNNs process input data step by step, where each step corresponds
to a time step in the sequence. They maintain hidden states that capture information
from previous time steps [24]. RNNs have recurrent connections that allow informa‑
tion to flow from one time step to the next. This enables the network to maintain con‑
text and capture sequential dependencies [25]. The hidden state of an RNN at each
time step stores information from the current input and the previous hidden state. It
serves as the network’s memory. RNNs can suffer from vanishing gradient problems,
where gradients become extremely small as they are propagated back in time during
training. This can hinder long‑range dependencies.

1.6.1 Types of RNNs
1. Simple RNN: The basic RNN architecture processes sequences step by
step, but it suffers from the vanishing gradient problem, making it difficult
to capture long‑term dependencies.
2. Long Short‑Term Memory (LSTM): LSTM networks address the vanish‑
ing gradient problem by incorporating specialized memory cells that can
store and access information over long sequences.
12 Data-Centric Artificial Intelligence for Multidisciplinary Applications

3. Gated Recurrent Unit (GRU): Similar to LSTM, GRU also mitigates the
vanishing gradient problem using gating mechanisms. It has a simplified
structure with fewer parameters.

1.6.2 Applications of RNNs
RNNs are widely used in various applications involving sequential data:

• NLP: Language modelling, machine translation, sentiment analysis, and

text generation.
• Speech Recognition: Converting spoken language into text.
• Time Series Analysis: Predicting future values in financial markets,
weather forecasting, and stock prices.
• Music Generation: Creating new music sequences.
• Video Analysis: Action recognition and video captioning.

While RNNs are powerful for handling sequential data, they also have limitations,
such as difficulty in capturing very long‑range dependencies. More advanced archi‑
tectures like Transformers have emerged to address some of these limitations and
have gained popularity in various applications as well [26].
The above‑discussed AI technologies have some ethical and legal implica‑
tions in modern day‑to‑day life decision‑making processes. These implications are
­discussed below.

1.7 UNDERSTANDING THE ETHICAL

IMPLICATIONS OF AI TECHNOLOGIES
Understanding the ethical implications of AI technologies is crucial as AI systems
become increasingly integrated into our lives and decision‑making processes [27].
AI technologies have the potential to bring about transformative benefits, but they
also raise ethical concerns that need to be carefully considered and addressed. AI
systems are susceptible to inheriting biases embedded within their training data,
culminating in outcomes that are unjust or discriminatory. This is particularly evi‑
dent in domains such as recruitment, lending, and criminal justice. Ensuring fair‑
ness and reducing bias require careful data collection, unbiased model design, and
ongoing monitoring. Many AI models, especially deep learning models, can be dif‑
ficult to interpret and explain. This lack of transparency can lead to distrust and
hinder accountability. Efforts are being made to develop methods for explaining AI
decisions, providing transparency, and making models more interpretable [28]. AI
systems frequently engage in the processing of extensive personal data, giving rise
to apprehensions about potential breaches of privacy and unauthorized data access.
Privacy‑preserving techniques and strict data‑handling practices are necessary to
protect individuals’ sensitive information. The automation of certain tasks by AI
could lead to job displacement in certain industries, potentially affecting workers
and local economies. Addressing the economic and social impact of automation may
require upskilling and reskilling programmes. Determining responsibility when AI
Advancements in Data-Centric AI 13

systems make errors or harmful decisions can be challenging. There may be ­questions
about whether the responsibility lies with the developers, users, or the technology
itself [29]. Also, clear guidelines for accountability and liability are essential, espe‑
cially in safety‑critical applications. AI systems become more autonomous, ques‑
tions arise about who has control over their decisions and actions. Ensuring human
oversight, intervention mechanisms, and fail‑safes are crucial to prevent unintended
consequences. AI systems can be vulnerable to adversarial attacks, where malicious
actors manipulate inputs to deceive the system. Robustness testing and security
measures are necessary to safeguard AI systems from such attacks. Complex AI
systems can exhibit behaviour that was not explicitly programmed, leading to unex‑
pected outcomes. Comprehensive testing and validation procedures are essential to
identify and mitigate unintended consequences. AI technologies have the potential
to reshape society in profound ways, impacting economies, job markets, and social
norms. Long‑term ethical considerations and societal implications should guide the
development and deployment of AI technologies [30]. Ensuring that AI technologies
benefit all of humanity, regardless of geographic location or socioeconomic status,
is a significant ethical concern. Efforts should be made to bridge the digital divide
and prevent exacerbating existing inequalities. Addressing these ethical implications
requires collaboration between policymakers, researchers, industry stakeholders,
and society as a whole. Ethical frameworks, regulations, and guidelines are being
developed to ensure that AI technologies are developed and deployed in ways that
prioritize human well‑being, fairness, and accountability.

1.8 PRIVACY AND DATA PROTECTION IN AI‑DRIVEN SYSTEMS

Ensuring privacy and safeguarding data are pivotal considerations in the design and
implementation of AI‑driven systems. With the increasing prevalence and advance‑
ment of AI technologies, substantial data access is often imperative for optimal func‑
tionality. However, this access must be coupled with robust mechanisms that ensure
the confidentiality and security of individuals’ sensitive information. The following
key aspects merit attention:

• Prudent Data Collection: Gather only the essential data required for the
intended purpose of the AI system. Restricting data acquisition mitigates
the potential for privacy breaches and unauthorized utilization [31].
• Knowledgeable Consent: Prior to collecting and utilizing personal data,
secure informed and explicit consent from individuals. Individuals should
have a comprehensive understanding of data usage and the opportunity to
opt out [32].
• Anonymization and De‑identification: Modify or eliminate personally
identifiable information from data to avert direct individual identification.
Notably, it’s vital to acknowledge that anonymization might not always
guarantee privacy, considering potential re‑identification attacks.
• Robust Data Security: Enforce stringent security measures to shield amassed
data from unauthorized access, breaches, and cyber threats. Encryption,
secure storage, and access controls assume paramount importance.
14 Data-Centric Artificial Intelligence for Multidisciplinary Applications

• Purpose Constraint: Confine the utilization of collected data to the spe‑

cific purposes for which consent was granted. Utilizing data for unrelated
objectives should be avoided unless supplementary consent is obtained.
• User Oversight and Transparency: Furnish users with lucid, easily acces‑
sible details regarding data application. Empower users to manage their data
and extend choices for data erasure.
• Data Retention Guidelines: Define and abide by data retention durations
that align with the rationale behind data collection. Protracted data reten‑
tion beyond necessity should be averted.
• Cross‑Border Data Transfer: Be mindful of regulations and laws regard‑
ing the transfer of data across borders. Different jurisdictions may have
varying requirements for data protection.
• Third‑Party Data Sharing: If sharing data with third parties (such as
cloud providers), ensure they adhere to strict data protection and privacy
standards. Contracts should outline data handling practices.
• Bias and Fairness: Be vigilant about potential biases in AI models and
data. Biased models could lead to discriminatory outcomes, impacting pri‑
vacy and fairness.
• Regulatory Compliance: Take the initiative to acquaint yourself with per‑
tinent data protection regulations, such as Europe’s General Data Protection
Regulation (GDPR) or the United States’ California Consumer Privacy Act
(CCPA). It is crucial to uphold compliance with these regulations to ensure
responsible and ethical data handling practices.
• Ethical Data Use: Develop ethical guidelines and policies that govern the
use of data in AI systems. Ensure that data usage aligns with ethical prin‑
ciples and societal norms.
• Regular Audits and Assessments: Periodically assess data privacy prac‑
tices, conduct privacy impact assessments, and perform audits to ensure
ongoing compliance and identify areas for improvement.
Balancing the benefits of AI‑driven systems with the protection of indi‑
viduals’ privacy is a complex but essential task [33]. Striving for transpar‑
ency, accountability, and ethical use of data is fundamental to building trust
and ensuring that AI technologies are developed and deployed responsibly.

The roles and responsibilities of data‑centric AI in various applications are given in

the following section.

1.9 DATA‑CENTRIC AI IN FINANCE

Data‑centric AI refers to an approach that places data at the core of AI applications,
focusing on the collection, processing, analysis, and interpretation of data to derive
valuable insights and make informed decisions.
Financial institutions gather data from various sources, including customer trans‑
actions, market data, social media, and economic indicators. Data‑centric AI involves
integrating and aggregating this diverse data efficiently to create a comprehensive
and unified dataset. Since financial data can be noisy and incomplete, preprocessing
Advancements in Data-Centric AI 15

and cleaning are essential steps in data‑centric AI [34]. Advanced algorithms are
used to clean and transform data, ensuring its quality and consistency. Data‑centric
AI in finance heavily relies on ML and AI algorithms to analyse the data, discover
patterns, and make predictions. Algorithms such as regression, classification, clus‑
tering, and deep learning models are commonly used. Financial institutions utilize
data‑centric AI to assess and manage risks associated with investments, loans, and
other financial products. By analysing historical data and market trends, AI mod‑
els can identify potential risks and provide insights to improve risk management
strategies. Data‑centric AI enables financial institutions to offer personalized ser‑
vices and recommendations to their customers. By analysing customer behaviour
and preferences, AI models can suggest tailored financial products and services that
meet individual needs. AI‑powered fraud detection systems are increasingly used
in finance to identify suspicious activities and transactions [35]. These systems can
detect anomalies in real‑time and prevent fraudulent activities, protecting both cus‑
tomers and financial institutions. Data‑centric AI plays a crucial role in ensuring
regulatory compliance for financial institutions. AI models can analyse vast amounts
of data to identify any non‑compliance issues and assist in meeting reporting require‑
ments. Data‑centric AI can analyse market trends, news, and other economic factors
to inform trading decisions. This approach is commonly used in algorithmic trading
and quantitative finance [36].

1.10 AI FOR FRAUD DETECTION

AI‑powered fraud detection systems use ML algorithms to analyse large volumes of
data and identify patterns that may indicate fraudulent activities [37]. These systems
can detect both known and unknown fraud patterns, providing real‑time alerts to
prevent potential financial losses. Key features of AI for fraud detection include:

• Anomaly Detection: AI algorithms can identify unusual behaviour or

transactions that deviate from normal patterns, helping to detect fraudulent
activities.
• Behavioural Analysis: By analysing historical customer behaviour, AI can
recognize changes or deviations in behaviour that might indicate fraud [38].
• Network Analysis: AI can assess relationships between various entities, such
as customers, merchants, and accounts, to identify potential fraud networks.
• ML Models: Supervised and unsupervised learning techniques are com‑
monly used to train models for fraud detection. These models continuously
learn from new data to improve accuracy.
• AI for Risk Assessment: AI finds application in financial risk assessment,
enhancing comprehension and management of diverse risk categories such
as credit, market, operational, and liquidity risks. Employing AI techniques
empowers financial entities to make judicious choices and allocate resources
adeptly. Notable applications encompass:
• Credit Risk Evaluation: AI models analyse a borrower’s credit history,
financial data, and other relevant factors to assess the risk associated with
extending credit.
16 Data-Centric Artificial Intelligence for Multidisciplinary Applications

• Market Risk Prediction: AI algorithms analyse market data and historical

trends to predict potential fluctuations and market movements.
• Operational Risk Management: AI can be employed to identify potential
operational risks and provide insights into improving internal processes.
• Stress Testing: AI‑powered stress testing models assess how financial
institutions’ portfolios and balance sheets would perform under adverse
scenarios.
• AI for Algorithmic Trading: Algorithmic trading involves using pre‑defined
rules and AI‑powered algorithms to execute trades at high speeds and fre‑
quencies. AI helps traders make data‑driven decisions, identify patterns in
market data, and execute trades more efficiently. Key aspects of AI in algo‑
rithmic trading include:
• Market Analysis: AI models analyse large volumes of historical and
real‑time market data to identify trading opportunities and patterns.
• NLP: AI can process and understand news articles and other textual data to
assess their impact on the financial markets.
• RL: AI agents can use RL techniques to learn and adapt trading strategies
based on feedback from the market [39].
• Risk Management: AI models can dynamically adjust trading strategies to
manage risk exposure and avoid large losses.

1.11 ENSURING FAIRNESS AND TRANSPARENCY

IN FINANCIAL AI APPLICATIONS
Ensuring fairness and transparency in financial AI applications is crucial to building
trust, preventing bias, and maintaining ethical standards in the financial industry.
Here are some key considerations to achieve fairness and transparency:

1. Data Quality and Bias Mitigation

• Start with high‑quality and diverse data to avoid bias in AI models.
Ensure that the data used for training is representative of the entire pop‑
ulation and includes samples from different demographics.
• Use techniques such as data preprocessing, augmentation, and balanc‑
ing to mitigate bias and address data imbalances [40].
2. Algorithmic Fairness
• Choose AI algorithms that are known for their fairness and interpret‑
ability. Certain algorithms may inherently introduce biases, so it’s
essential to evaluate their impacts on different groups.
• Implement fairness‑aware ML techniques that explicitly consider fair‑
ness constraints during model training and decision‑making processes.
3. Interpretability
• Utilize AI models that can provide explanations for their decisions.
Explainable AI helps users understand the factors that influenced a spe‑
cific outcome, making it easier to detect and address potential biases.
Advancements in Data-Centric AI 17

• Choose interpretable models, such as decision trees or linear regression,

whenever possible, as they offer clearer insights into how decisions are
made.
4. Regulatory Compliance
• Maintain awareness regarding pertinent financial regulations and data
protection legislation. Adherence to laws such as GDPR or CCPA is of
paramount importance when dealing with sensitive financial informa‑
tion [41].
• Ensure that AI applications meet the necessary legal and regulatory
requirements, including anti‑discrimination laws.
5. Model Validation and Testing
• Thoroughly evaluate AI models for fairness before deployment. Use
fairness metrics to assess the impact on different groups and demo‑
graphics [42].
• Conduct rigorous testing, including stress testing, to identify potential
biases and weaknesses in the AI system.
6. Human Oversight
• Involve human experts in the development and monitoring of AI appli‑
cations. Humans can provide insights into potential biases that AI might
not detect.
• Establish a governance framework that includes ethical committees or
review boards to oversee AI usage and decision‑making.
7. Continuous Monitoring and Feedback
• Sustain a perpetual vigil over the performance of AI models in live
operation to identify potential drift or bias that may emerge over time.
• Instigate a culture of eliciting feedback from end‑users and stakeholders,
a process that unveils and rectifies fairness or transparency concerns.
8. Transparency in Decision‑Making
• Communicate openly about the use of AI in financial applications with
customers and stakeholders.
• Be transparent about the factors that AI models consider in deci‑
sion‑making processes.

There may have been significant developments in AI. Key emerging technologies and
trends in AI up to the point are discussed in Section 1.12.

1.12 EMERGING TECHNOLOGIES IN AI

Edge AI and Internet of Things (IoT) applications are two intertwined emerging
technologies that bear considerable potential for shaping the trajectory of AI in the
future. Edge AI signifies the integration of AI algorithms onto edge devices, encom‑
passing smartphones, sensors, cameras, and various IoT devices, in contrast to rely‑
ing on centralized cloud servers for computation [43]. This approach, processing data
directly on the device, imparts a range of benefits:
18 Data-Centric Artificial Intelligence for Multidisciplinary Applications

1.12.1 Edge AI and Internet of Things Applications

• Real‑time Processing: Edge AI enables faster decision‑making as data is
processed on the device without the need for round‑trips to cloud servers,
reducing latency.
• Privacy and Security: Since data remains on the device and is not
transmitted to external servers, Edge AI can enhance data privacy and
security.
• Bandwidth Efficiency: Edge AI diminishes the volume of data necessitat‑
ing transmission to cloud servers, thereby optimizing network bandwidth
and curbing expenses associated with data transfers [44].
• Offline Functionality: Edge AI allows devices to perform AI tasks even
when they are disconnected from the internet, making them more resilient
in remote or low‑connectivity environments.
• Scalability: Distributing AI computation across edge devices can scale
effectively without overburdening centralized cloud infrastructure.
• IoT Applications: The IoT constitutes a network encompassing intercon‑
nected devices, imbued with sensors, software, and other technological
components that empower them to gather and share data via the internet.
The amalgamation of AI with IoT devices ushers in a myriad of prospects
for a wide array of applications.
• Smart Home Automation: AI‑powered IoT devices can learn and adapt
to users’ preferences, optimizing energy consumption and enhancing home
security.
• Industrial IoT: The amalgamation of AI and IoT within industrial environ‑
ments facilitates predictive maintenance, process enhancement, and auto‑
mation. This synergy translates to heightened efficiency and minimized
downtime.
• Healthcare IoT: IoT devices with AI capabilities can monitor patients’
health conditions remotely, facilitate early disease detection, and improve
healthcare outcomes [45].
• Smart Cities: AI‑driven IoT infrastructure can improve urban planning,
traffic management, waste management, and public safety.
• Environmental Monitoring: IoT sensors with AI analytics can track
­environmental parameters, such as air quality and water quality, for better
environmental management.
• Agriculture: AI‑powered IoT devices assist in precision agriculture by opti‑
mizing irrigation, monitoring crop health, and predicting weather patterns.
• Wearable Devices: Wearable IoT devices with AI functionalities can moni‑
tor fitness, health metrics, and provide personalized insights.

Both Edge AI and IoT applications are evolving rapidly, and their combined poten‑
tial has significant implications for various industries, ranging from consumer elec‑
tronics to healthcare, manufacturing, transportation, and beyond. The integration of
AI capabilities at the edge with IoT devices is expected to drive further innovation
Advancements in Data-Centric AI 19

and enhance the efficiency, intelligence, and capabilities of connected devices and
systems.
Quantum computing has the potential to revolutionize AI by significantly speed‑
ing up complex computations, enabling researchers to tackle more intricate AI
problems.

1.13 QUANTUM COMPUTING AND ITS

POTENTIAL IMPACT ON AI
Quantum computing possesses the potential to usher in a revolutionary transfor‑
mation within the realm of AI, offering unparalleled computational prowess and
resolving intricate problems that currently remain beyond the reach of classical com‑
puters. Rooted in the principles of quantum mechanics, quantum computing har‑
nesses quantum bits (qubits) that can exist in a multitude of states concurrently. This
attribute facilitates parallel processing and exponential scalability, thereby present‑
ing new horizons for computational capabilities. Quantum computing’s ability to
perform complex calculations exponentially faster than classical computers could
significantly speed up AI algorithms. Tasks like optimization, search, and pattern
recognition could be accelerated, leading to more efficient AI systems [46]. Quantum
computing can enhance ML algorithms by enabling the exploration of larger fea‑
ture spaces and optimizing models with higher‑dimensional data. Quantum ML
algorithms may outperform classical counterparts in certain scenarios. Quantum
computing can efficiently analyse vast amounts of data, making it valuable for pro‑
cessing and extracting insights from big data, which is essential for AI applications.
Quantum computing’s parallel processing capabilities make it ideal for solving com‑
plex optimization problems. AI applications that rely on optimization, such as port‑
folio optimization in finance or resource allocation in logistics, could benefit from
quantum‑enhanced algorithms. Researchers are exploring the concept of quantum
neural networks, which leverage quantum entanglement and superposition to process
information. Quantum neural networks may be more adept at processing complex
data and making predictions in specific scenarios [47]. Quantum computing could
improve NLP tasks like language translation, sentiment analysis, and document
summarization by processing large amounts of text data more efficiently. Quantum
computers could simulate complex quantum systems, which would be invaluable
for quantum‑inspired AI research, quantum chemistry, and drug discovery. Despite
these promising prospects, it’s important to note that practical quantum computing is
still in its early stages. Building reliable, error‑corrected quantum computers remains
a considerable challenge due to quantum decoherence and noise. Currently, most
quantum processors are noisy and limited in qubit count, making them unsuitable for
many large‑scale AI applications. As quantum computing technology advances and
matures, it is expected to have a transformative impact on AI [48]. Researchers are
actively exploring quantum algorithms, hybrid quantum‑classical approaches, and
the integration of quantum computing with classical AI techniques to harness the full
potential of quantum computing for AI applications.
20 Data-Centric Artificial Intelligence for Multidisciplinary Applications

1.14 AI GOVERNANCE AND REGULATION

AI governance and regulation refer to the policies, rules, and frameworks estab‑
lished by governments, organizations, and international bodies to guide the
responsible development, deployment, and use of AI technologies. Given the
potential societal impact of AI, it is essential to ensure that these technologies
are developed and used in an ethical, transparent, and accountable manner [49].
Many AI governance initiatives are based on ethical principles that promote fair‑
ness, transparency, accountability, privacy, and human rights. These principles
serve as guiding values to ensure that AI systems benefit society while minimizing
potential risks and harm. Regulations may require AI developers to ensure trans‑
parency in how their systems operate and provide explanations for the decisions
made by AI algorithms, especially in critical domains like healthcare and finance.
Regulations often address data protection concerns related to AI applications,
ensuring that personal and sensitive data are handled securely and with user con‑
sent. Governance frameworks strive to mitigate algorithmic bias and ensure that AI
systems do not discriminate against individuals or groups based on attributes like
race, gender, or religion. Clear guidelines on accountability and liability are cru‑
cial to determine responsibility when AI systems cause harm or make erroneous
decisions [50]. Some regulations focus on high‑risk AI applications, such as critical
infrastructure, healthcare, and autonomous vehicles, requiring additional scrutiny,
testing, and certification. In an interconnected world, international cooperation
is necessary to address cross‑border data flows, privacy concerns, and regulatory
harmonization. Effective governance involves engaging with the public, AI devel‑
opers, researchers, and other stakeholders to understand concerns, assess risks, and
foster trust. Regulations might address the impact of AI on the workforce, promot‑
ing responsible workforce management, and addressing potential job displacement.
Governments may invest in AI research, promote collaboration between academia
and industry, and encourage responsible AI innovation. Developing AI standards
and certification processes can help ensure compliance with regulations and foster
a culture of responsible AI development. In light of the worldwide reach of AI
technologies, fostering international collaboration among governments and organi‑
zations is imperative to establish uniform and effective governance principles. The
landscape of AI governance and regulation is dynamically evolving, with diverse
countries and organizations adopting varied strategies to tackle the ethical, legal,
and societal quandaries posed by AI [51]. As AI technologies advance, maintaining
an equilibrium between fostering innovation and ensuring responsible AI develop‑
ment and utilization for the greater good of humanity remains pivotal.

1.15 STRATEGIES FOR RESPONSIBLE AI

IMPLEMENTATION AND OVERSIGHT
Enforcing responsible AI implementation and vigilant oversight is vital to ensure
that AI systems are conceived, rolled out, and employed in a manner that aligns
with ethical tenets, legal requisites, and societal values. Following are some
strategies:
Advancements in Data-Centric AI 21

• Incorporate ethical considerations into the inception of AI systems’ design

and development.
• Engage multidisciplinary teams encompassing ethicists, data scientists, and
domain experts to pinpoint potential biases, equity concerns, and ethical
ramifications [52].
• Affirm transparency and comprehensibility in AI systems, furnishing
insights into their decision‑making processes.
• Leverage interpretable AI methodologies to expound AI predictions and
actions, especially in critical domains like healthcare and finance.
• Assign primacy to top‑tier, diverse data to preclude bias in AI models. Employ
data preprocessing techniques to curb bias and tackle data imbalances.
• Undertake periodic audits and surveillance of data to detect potential biases [53].
• Incorporate human oversight and intervention mechanisms in AI systems.
• Establish feedback loops where users can report issues and provide feed‑
back to improve system performance and mitigate biases.
• Thoroughly test AI algorithms before deployment to identify and address poten‑
tial biases and shortcomings, and conduct robust validation processes including
evaluating AI performance on diverse datasets and simulated scenarios.
• Continuously monitor AI systems in real‑world settings to detect any unin‑
tended consequences or changes in behaviour. Implement auditing mecha‑
nisms to assess AI performance regularly.
• Establish review boards or ethics committees to assess and evaluate AI proj‑
ects, especially in sensitive domains like healthcare and criminal justice.
• Undertake comprehensive assessments to comprehend the potential ramifi‑
cations of introducing AI systems in diverse contexts, taking into account
social, ethical, and legal ramifications.
• Craft and adhere to lucid guidelines and policies governing the ethical
deployment of AI within organizational frameworks. These directives
should encompass aspects such as data privacy, security, impartiality, and
ethical considerations. Stay attuned to pertinent AI regulations and ensure
alignment with regulations on data protection and privacy.
• Acquire a deep understanding of the regulatory milieu in regions where AI
systems will be implemented.
• Foster awareness among end‑users regarding the capabilities and limita‑
tions of AI systems, furnishing unambiguous details about data usage.
• Secure informed consent before collecting and processing personal data.
Implement robust risk management strategies to address potential AI‑related
risks, including financial, reputational, and operational risks.
• Establish mechanisms to report AI‑related incidents or concerns, and be
transparent with stakeholders about actions taken to address such issues.

1.16 COLLABORATION AND KNOWLEDGE SHARING

Engage in knowledge sharing and collaboration with other organizations and stake‑
holders to learn from best practices and collectively address challenges related to
responsible AI deployment. By following these strategies, organizations can deploy
22 Data-Centric Artificial Intelligence for Multidisciplinary Applications

AI systems responsibly and ensure that AI technologies contribute positively to soci‑

ety while minimizing potential risks and negative impacts. Effective monitoring will
allow for continuous improvement and ensure that AI systems remain aligned with
ethical and societal standards over time.

1.17 FUTURE OF DATA‑CENTRIC AI

The future of data‑centric AI is expected to be shaped by several trends and develop‑
ments in AI technologies while simultaneously facing ethical challenges that need to
be addressed responsibly.

1.17.1 Trends and Future Developments in AI Technologies

Deep learning will continue to evolve, enabling more complex and sophisticated AI
applications. Researchers may explore novel architectures and optimization tech‑
niques to improve the performance of deep neural networks. Transfer learning and
few‑shot learning approaches will become more prevalent, allowing AI models to
learn from one task and apply that knowledge to new tasks with limited data. AI sys‑
tems will integrate information from multiple sources, such as text, images, speech,
and sensor data, to create more comprehensive and context‑aware models. XAI will
gain prominence to address the black‑box nature of some AI models, making AI
systems more transparent and interpretable, especially in critical applications like
healthcare and finance. As quantum computing advances, quantum‑inspired AI algo‑
rithms will emerge, leveraging quantum principles to accelerate AI computations.
Edge AI is poised to maintain its increasing significance, facilitating AI computa‑
tions on edge devices such as IoT devices. This trend brings about reduced latency,
heightened privacy, and enhanced efficiency. Generative AI models will evolve to
create more realistic and creative outputs, impacting domains like art, design, and
content generation. AI will play a vital role in drug discovery, personalized medicine,
and disease prediction, contributing to advancements in biotechnology and health‑
care. AI will continue to be integrated into autonomous vehicles, drones, and robots,
enabling safer and more intelligent automation.

1.18 ETHICAL CHALLENGES IN EVOLVING AI LANDSCAPES

1. Addressing Bias and Ensuring Fairness: AI systems have the potential
to adopt and magnify biases ingrained in the data, resulting in inequitable
results. The imperative to maintain fairness and counteract bias will persist
as central concerns in AI advancement
2. Safeguarding Privacy and Data: Given AI’s dependence on extensive
data, safeguarding user privacy and confidential information assumes para‑
mount significance, mandating the implementation of robust data protection
protocols.
3. AI in Employment: The impact of AI on the workforce raises ethical
considerations about job displacement and the need for reskilling and up‑­
skilling programs.
Advancements in Data-Centric AI 23

4. Autonomous Systems and Safety: As AI powers autonomous systems,

ensuring their safety and accountability becomes essential to prevent poten‑
tial harm.
5. Weaponization of AI: The misuse of AI technologies for malicious pur‑
poses, such as in cyber‑attacks or autonomous weapons, poses significant
ethical challenges.
6. Transparency and Explainability: AI systems’ lack of transparency and
explainability can lead to a loss of trust, making it essential to address these
issues for critical applications.
7. Social Impact: AI’s influence on society, from misinformation to echo
chambers in social media, requires ethical considerations and responsible
design.
8. Regulatory and Legal Challenges: Developing and enforcing AI
­regulations that strike the right balance between innovation and account‑
ability remains a challenge for policymakers. Tackling these ethical dilem‑
mas demands a collective endeavour involving researchers, policymakers,
enterprises, and the general public. Nurturing responsible AI advancement,
continuous vigilance, and unwavering commitment to ethical directives
will stand as paramount. This approach is pivotal to guarantee that AI
­technologies serve humanity’s welfare while upholding ethical benchmarks
and societal principles.

REFERENCES
1. Abdi, H., and Williams, L. J. Principal component analysis. Wiley Interdisciplinary
Reviews: Computational Statistics 2, 4 (2010), 433–459.
2. Ahsan, M. M., Mahmud, M. P., Saha, P. K., Gupta, K. D., and Siddiqe, Z. Effect of data
scaling methods on machine learning algorithms and model performance. Technologies
9, 3 (2021), 52
3. Agarwal, A., Dahleh, M., & Sarkar, T. A marketplace for data: An algorithmic solution.
In Proceedings of the 2019 ACM Conference on Economics and Computation (2019)
(pp. 701–726).
4. Ali, P. J. M., Faraj, R. H., Koya, E., Ali, P. J. M., and Faraj, R. H. Data normalization and
standardization: a technical report. Machine Learning Technical Reports 1, 1 (2014), 1–6.
5. Armbrust, M., Ghodsi, A., Xin, R., & Zaharia, M. (2021, January). Lakehouse: a
new generation of open platforms that unify data warehousing and advanced analyt‑
ics. In Proceedings of CIDR (Vol. 8).6. Arocena, P. C., Glavic, B., Mecca, G., Miller,
R. J., Papotti, P., and Santoro, D. Benchmarking data curation systems. IEEE Data
Engineering Bulletin 39, 2 (2016), 47–62.
7. Aroyo, L., Lease, M., Paritosh, P., & Schaekermann, M. (2022). Data excellence for AI:
why should you care?. Interactions, 29, 2, 66–69.
8. Azhagusundari, B., Thanamani, A. S. et al. Feature selection based on information
gain. International Journal of Innovative Technology and Exploring Engineering
(IJITEE) 2, 2 (2013), 18–21.
9. Azizzadenesheli, K., Liu, A., Yang, F., & Anandkumar, A. (2019). Regularized learning
for domain adaptation under label shifts. arXiv preprint arXiv:1903.09734.10. Anand,
D., and Kumar, A. IoT‑based automated healthcare system. In: 5th International
Conference on Computing Methodologies and Communication (ICCMC), Erode,
India (2021).
24 Data-Centric Artificial Intelligence for Multidisciplinary Applications

11. Pao, L. Y., and Frei, C. W. A comparison of parallel and sequential implementations of
a multisensor multitarget tracking algorithm. In: Proceedings of the American Control
Conference, vol. 3, Seattle, WA, June (1995), pp. 1683–1687.
12. D’Mello, S., and Graesser, A. (2012). Dynamics of affective states during complex
learning. Learning and Instruction 22(2), 145–157
13. Baik, C., Jagadish, H. V., & Li, Y. (2019, April). Bridging the semantic gap with SQL
query logs in natural language interfaces to databases. In 2019 IEEE 35th International
Conference on Data Engineering (ICDE) (pp. 374–385). IEEE.
14. Barandas, M., Folgado, D., Fernandes, L., Santos, S., Abreu, M., Bota, P., Liu, H.,
Schultz, T. and Gamboa, H., 2020. TSFEL: Time series feature extraction library.
SoftwareX, 11, p.100456.
15. Basu, A., & Blanning, R. W. (1995, January). Discovering implicit integrity constraints
in rule bases using metagraphs. In Proceedings of the Twenty‑Eighth Annual Hawaii
International Conference on System Sciences (Vol. 3, pp. 321–329). IEEE.
16. Batini, C., Cappiello, C., Francalanci, C., and Maurino, A. Methodologies for data
­quality assessment and improvement. ACM Computing Surveys (CSUR) 41, 3 (2009),
1–52.
17. Baylor, D., Breck, E., Cheng, H. T., Fiedel, N., Foo, C. Y., Haque, Z., ... & Zinkevich, M.
(2017, August). Tfx: A tensorflow‑based production‑scale machine learning platform.
In: Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining (pp. 1387–1395)
18. Bertini, E., & Lalanne, D. (2009, June). Surveying the complementary role of auto‑
matic data analysis and visualization in knowledge discovery. In: Proceedings of the
ACM SIGKDD Workshop on Visual Analytics and Knowledge Discovery: Integrating
Automated Analysis with Interactive Exploration (pp. 12–20).
19. Yin, L., Lin, S., Sun, Z., Wang, S., Li, R., & He, Y. (2024). PriMonitor: An adaptive
­tuning privacy-preserving approach for multimodal emotion detection. World Wide
Web 27(2), 1–28.
20. D’Mello, S., Kory, J., and Dieterle, E. Affective computing: Challenges, techniques,
and evaluation. In: S. D’Mello, R. A. Calvo, and J. Gratch (Eds.), Handbook of Affective
Computing, Oxford University Press (2015), pp. 401–414.
21. Li, Y., Chen, Z., Zha, D., Zhou, K., Jin, H., Chen, H., and Hu, X. Automated anomaly
detection via curiosity‑guided search and self‑imitation learning. IEEE Transactions on
Neural Networks and Learning Systems 33, 6 (2021), 2365–2377.
22. Li, Y., Chen, Z., Zha, D., Zhou, K., Jin, H., Chen, H., & Hu, X. (2021, April). Autood:
Neural architecture search for outlier detection. In: 2021 IEEE 37th International
Conference on Data Engineering (ICDE) (pp. 2117–2122). IEEE.
23. Liu, P., Yuan, W., Fu, J., Jiang, Z., Hayashi, H., and Neubig, G. Pre‑train, prompt, and
predict: a systematic survey of prompting methods in natural language processing.
ACM Computing Surveys 55, 9 (2023), 1–35.
24. Li, Y., Zha, D., Venugopal, P., Zou, N., and Hu, X. PyODDS: An end‑to‑end outlier
detection system with automated machine learning. In: Companion Proceedings of the
Web Conference 2020 (2020).
25. Liu, Z., Chen, S., Zhou, K., Zha, D., Huang, X., and Hu, X. RSC: Accelerating graph
neural networks training via randomized sparse computations. In: Proceedings of the
40th International Conference on Machine Learning, PMLR 202:21951‑21968, 2023.
26. Liu, Z., Wei, P., Jiang, J., Cao, W., Bian, J., and Chang, Y. MESA: boost ensemble imbal‑
anced learning with meta‑sampler. In: Advances in Neural Information Processing
Systems 33 (NeurIPS 2020) (2020).
27. Lucic, Ana, Harrie Oosterhuis, Hinda Haned and M. de Rijke. FOCUS: Flexible
Optimizable Counterfactual Explanations for Tree Ensembles. AAAI Conference on
Artificial Intelligence (2019).
Advancements in Data-Centric AI 25

28. Luo, Y., Qin, X., Tang, N., and Li, G. DeepEye: Towards automatic data visualiza‑
tion. In: 2018 IEEE 34th International Conference on Data Engineering (ICDE), IEEE
(2018), pp. 101–112.
29. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., and Vladu, A. Towards deep learning
models resistant to adversarial attacks. In: 6th International Conference on Learning
Representations, ICLR 2018 ‑ Conference Track Proceedings. (2018)8.
30. Marcus, Ryan, Andreas Kipf, Alexander van Renen, Mihail Stoian, Sanchit Misra,
Alfons Kemper, Thomas Neumann and Tim Kraska. Benchmarking learned indexes.
In: Proceedings of the VLDB Endowment 14 (2020): 1–13.
31. M. Feurer, A. Klein, K. Eggensperger, J. Springenberg, M. Blum, and F. Hutter. Ecient
and robust automated machine learning. Advances in neural information processing
systems, 28, 2015.
32. Moosavi‑Dezfooli, S., Fawzi, A., & Frossard, P. DeepFool: A Simple and Accurate
Method to Fool Deep Neural Networks. 2016 IEEE Conference on Computer Vision
and Pattern Recognition (CVPR), 2574–2582. (2016).
33. Mazumder, M., Banbury, C.R., Yao, X., Karlavs, B., Rojas, W.G., Diamos, S., Diamos, G.F.,
He, L., Kiela, D., Jurado, D., Kanter, D., Mosquera, R., Ciro, J., Aroyo, L., Acun, B., Eyuboglu,
S., Ghorbani, A., Goodman, E.D., Kane, T., Kirkpatrick, C.R., Kuo, T., Mueller, J.W., Thrush,
T., Vanschoren, J., Warren, M.J., Williams, A., Yeung, S., Ardalani, N., Paritosh, P.K., Zhang,
C., Zou, J.Y., Wu, C., Coleman, C., Ng, A.Y., Mattson, P., & Reddi, V.J. DataPerf: Benchmarks
for Data‑Centric AI Development. ArXiv, abs/2207.10062. (2022).
34. Meduri, V., Popa, L., Sen, P., & Sarwat, M. (2020). A Comprehensive Benchmark
Framework for Active Learning Methods in Entity Matching. Proceedings of the 2020
ACM SIGMOD International Conference on Management of Data. (2020).
35. Mehrabi, N., Morstatter, F., Saxena, N.A., Lerman, K., & Galstyan, A.G. A Survey
on Bias and Fairness in Machine Learning. ACM Computing Surveys (CSUR), 54,
pp. 1– 35. (2019)
36. Meng, C., Trinh, L., Xu, N., Enouen, J., & Liu, Y. Interpretability and fairness evalua‑
tion of deep learning models on MIMIC‑IV dataset. Scientific Reports, 12. (2022).
37. Milutinovic, M., Schoenfeld, B., Martinez‑Garcia, D., Ray, S., Shah, S., and Yan, D. On
evaluation of autoML systems In: Computer Science (2020).
38. Mintz, M.D., Bills, S., Snow, R., & Jurafsky, D. Distant supervision for relation extrac‑
tion without labeled data. Annual Meeting of the Association for Computational
Linguistics. (2009).
39. Miotto, R., Wang, F., Wang, S., Jiang, X., and Dudley, J. T. Deep learning for healthcare:
review, opportunities and challenges. Briefings in Bioinformatics 19, 6 (2018).
40. Miranda, L. J. Towards data‑centric machine learning: a short review. Communications
of the ACM, 66, 8, pp. 84–92 10.1145/3571724 (2023)
41. Mirdita, M., Von Den Driesch, L., Galiez, C., Martin, M. J., Söding, J., and Steinegger,
M. Uniclust databases of clustered and deeply annotated protein sequences and align‑
ments. Nucleic Acids Research 45, D1 (2017), D170–D176.
42. Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D. &
Riedmiller, M. Playing atari with deep reinforcement learning. arXiv preprint
arXiv:1312.5602. (2013).
43. Nanni, L., Paci, M., Brahnam, S., and Lumini, A. Comparison of different image data
augmentation approaches. Journal of Imaging 7, 12 (2021).
44. Nargesian, F., Zhu, E., Pu, K. Q., and Miller, R. J. Table union search on open
data. In: Proceedings of the VLDB Endowment 11, 7 813–825 (2018). https://doi.
org/10.14778/3192965.3192973
45. Kim, J. S., Jin, H. et al., Location‑based social network data generation based on pat‑
terns of life. In: 2020 21st IEEE International Conference on Mobile Data Management
(MDM), IEEE (2020), pp. 158–167.
26 Data-Centric Artificial Intelligence for Multidisciplinary Applications

46. Kumar, A., Dabas, V., and Hooda, P., 2020. Collaboration Big Data. Internet of data: a
SWOT analysis. International Journal of Information Technology 12 (2020) 1159–1169.
47. Van Aken, D., Pavlo, A., Gordon, G. J., and Zhang, B. Automatic database management
system tuning through large‑scale machine learning. In SIGMOD ‘17: Proceedings
of the 2017 ACM International Conference on Management of Data May (2017)
1009–1024 https://doi.org/10.1145/3035918.3064029 (2017).
48. Venkatasubramanian, S., and Alfano, M., 2020. The philosophical basis of algorithmic
recourse. In Proceedings of the 2020 conference on fairness, accountability, and trans‑
parency, pp. 284–293.
49. Zha, D., Lai, K.‑H., Tan, Q., Ding, S., Zou, N., and Hu, X. B. Towards automated
­imbalanced learning with deep hierarchical reinforcement learning. In Proceedings of
the 31st ACM International Conference on Information & Knowledge Management.
2476–2485. (2022).
50. Pedrozo, W. G., Nievola, J. C., and Ribeiro, D. C. An adaptive approach for index tuning
with learning classifier systems on hybrid storage environments, In: Hybrid Artificial
Intelligent Systems: 13th International Conference, HAIS 2018, Proceedings, vol. 13,
Oviedo, Spain, June 20–22, 2018, Springer (2018), pp. 716–729.
51. Bodenheimer, T., Sinsky, C., and Froman, R. Improving Primary Care Access and
Continuity: A Framework and Quality Improvement Toolkit for Achieving the Institute
for Healthcare Improvement’s Triple Aim. Agency for Healthcare Research and Quality
(US) (2018).
52. Woolf, B. P. Building Intelligent Interactive Tutors: Student‑Centered Strategies for
Revolutionizing E‑Learning. Morgan Kaufmann (2010).
53. Lee, Y., Im, D., and Shim, J., 2019. Data labeling research for deep learning based fire
detection system. In: 2019 International Conference on Systems of Collaboration Big
Data. Internet of Things and Security (SysCoBIoTS), IEEE (2021), pp. 1–4.
 2 Emerging Development
and Challenges in
Data‑Centric AI
Chaitali Shewale

2.1 INTRODUCTION
Data are crucial for training models, assessing performance, and generating predic‑
tions in artificial intelligence (AI). The concept that the quality and quantity of data
have a direct impact on the efficacy of AI models is at the heart of data‑centric AI
(DCAI). This chapter aims to shed light on how the AI community is developing to
adapt to the changing landscape of data by examining the new trends and difficulties
in DCAI. In this fast‑growing world, AI has rapidly transformed much of our human
work into a machine world with one click. The main core of this transformation lies
in the paradigm shift toward DCAI, where the huge and unprecedented growth along
with the availability of data has become a challenge and also an advantage for us.
With this huge amount of data, there is a need for a mechanism to separate and dis‑
tribute data to make it used more efficiently and in a much better way. The concept of
DCAI will have a direct effect on data arrangement and data distribution [1]. We will
delve into the techniques and strategies employed in data preprocessing, augmenta‑
tion, and feature engineering that contribute to the enhancement of data quality and
subsequent efficacy of models.
Applications for DCAI are prevalent across a range of industries, highlighting
the critical role that data plays in forming and improving AI models. The use of
large medical imaging data by AI models for the accurate detection of diseases
like cancer and the analysis of electronic health records to identify patterns and
provide individualized therapies are two notable examples. By anticipating main‑
tenance requirements and minimizing downtime, predictive maintenance, based
on sensor data analysis, optimizes operations. Image processing and data analysis
help quality control find defects in manufactured goods and guarantee excellent
quality. AI models analyze transaction history to find anomalous tendencies sug‑
gestive of probable fraud in financial transactions. Energy consumption optimiza‑
tion increases the effectiveness of energy distribution and lowers costs, whereas
traffic management enhances the efficiency of transportation networks and traffic
flow through data analysis. Through the use of multilingual text data, AI models are
trained to improve the accuracy of language translation systems. Sentiment analy‑
sis uses a massive amount of text data from reviews, surveys, and social media to

DOI: 10.1201/9781003461500-3 27
28 Data-Centric Artificial Intelligence for Multidisciplinary Applications

evaluate public opinion on a variety of topics. Analyzing historical soil, crop, and
meteorological data helps predict crop yields, increasing farm production. Real‑time
decision‑making and obstacle identification are made possible for autonomous cars
through image and sensor processing. By making recommendations for goods or
services based on user preferences, recommender systems increase consumer happi‑
ness and revenue. In order to keep the right amount of inventory on hand, inventory
optimization ­analyzes sales data and demand forecasts. In finance, risk assessment is
evaluating and predicting investment risks by examining market and economic data.
These fields have undergone a revolution, thanks to DCAI, which emphasizes how
crucial it is to have high‑quality, diversified data for AI applications. These illustra‑
tions highlight how DCAI applications use a wealth of data to build models, forecast
the future, and streamline processes in a variety of industries, ultimately enhancing
productivity, accuracy, and judgment.

2.2 DATA‑CENTRIC AI (DCAI)

The development of AI models using machine learning (ML) generally assumes that
the dataset is reasonably clean and well‑curated (for example, pictures of dogs and
cats), and the task is to create the best model for this dataset. The ML study covers a
variety of models such as neural networks, decision trees, etc., training methods such
as regularization, optimization algorithms, loss functions, etc., and model/hyperpa‑
rameter selection methods including model ensembling [2]. This theory is known
In real‑world applications, the data are not fixed like it is assumed when develop‑
ing AI model. Many times we are allowed to alter the dataset to improve model‑
ing accuracy or even add more data when your budget permits. Experienced data
scientists are aware that examining and improving the data is a better investment
than fiddling with models, but this process can be time‑consuming for huge datasets.
The majority of data improvement has been done manually under the direction of
human intuition or knowledge. Alternatively, we can methodically design data to
create superior AI systems. This approach is referred to as DCAI [3]. DCAI uses AI
technologies to more systematically identify and address problems that frequently
afflict real‑world datasets, whereas manual exploratory data analysis is an important
initial step in understanding and enhancing any dataset. AI that is data‑centric can
take one of two shapes:

1. AI algorithms that comprehend data and enhance models with that knowl‑
edge. This is seen in curriculum learning, where ML models are initially
trained on “easy data”.
2. Data‑modifying AI algorithms that enhance AI models. This is exemplified
by confident learning, in which ML models are trained on a filtered dataset
with incorrectly identified data eliminated [4].

Both of the aforementioned examples use algorithms to analyze the outputs of trained
ML models to automatically estimate which data is simple or incorrectly categorized.
The foundation of DCAI is the collection of diverse, high‑quality data [5]. The
popularity of methods like active learning and federated learning has made it
Emerging Development and Challenges in Data‑Centric AI 29

possible to collect data from many sources effectively [6]. Additionally, data labeling
and annotation are essential to supervised learning. The process of annotating data
is being revolutionized by improvements in methodologies like weak supervision and
semi‑supervised learning, which make the procedure more effective and scalable.
Cleaning, standardization, and feature scaling are examples of preprocessing tech‑
niques used to get data ready for AI models. By broadening the dataset, data augmen‑
tation methods including rotation, translation, and generative approaches improve
model performance. In order for the AI model to extract reliable and generalized
patterns from the data, preprocessing and augmentation are necessary. The privacy
of user data is a major concern in DCAI. Data privacy must be protected throughout
the AI lifecycle due to rising data breaches and privacy laws. Emerging technolo‑
gies like differential privacy, federated learning and homomorphic encryption enable
AI progress while protecting individual privacy. As a result of its use in modify‑
ing AI models for certain applications, domain‑specific data is becoming more and
more important. A potent method to use domain‑specific data is transfer learning,
a technique where models learned on one task are adapted for another related one.
Pre‑trained models can be fine‑tuned to drastically cut down on training time and
data requirements, opening up AI to more specialized sectors. Biases found in train‑
ing data can make AI models continue to be unfair and discriminatory. A persistent
problem in DCAI is addressing biases in the data and assuring fairness. To reduce
biases and encourage justice in AI applications, strategies like adversarial debiasing
and fairness restrictions are being investigated. Despite improvements, there are still
problems in the field of DCAI. Managing biases and ensuring privacy, along with the
rising demand for large‑scale, high‑quality data, provide significant hurdles. To meet
the changing demands of DCAI, future initiatives include establishing standardized

FIGURE 2.1 AI development life cycle.

30 Data-Centric Artificial Intelligence for Multidisciplinary Applications

tools for data management, producing synthetic data for improved generalization,
and encouraging interdisciplinary collaboration.
DCAI is the discipline of systematically engineering data used to build an AI sys‑
tem [7]. Many researchers have defined DCAI in many different ways, some of them
as “the discipline of systematically engineering the data used to build an AI system”.
As shown in Figure 2.1, the prime approach in DCAI is to take data to center stage or
move the data to a centralized unit to make data accessible in the entire AI develop‑
ment lifecycle. Here, we can access data taken at all the stages of the AI lifecycle.
DCAI is based on the objective of providing the best dataset to feed a given ML
model, while model‑centric AI is based on the goal of producing the best model for
a given dataset [8]. One should do both in order to implement the best‑supervised
learning systems in practice. A DCAI includes the following activities:

1. Investigate the data, correct any major problems, and then change it so that
it is ML‑appropriate.
2. Utilizing the correctly prepared dataset, train a foundational ML model.
3. Make use of this model to help you enhance the dataset using the methods
covered in this class.
4. To obtain the optimal model, experiment with various modeling strategies
on the improved dataset

The greatest way to boost performance in real‑world ML applications is frequently

through the study of techniques to enhance datasets, which is what DCAI is study‑
ing. While competent data scientists have long used trial‑and‑error and intuition to
improve data, DCAI views data improvement as a structured engineering discipline.

2.2.1 Key Components of DCAI

1. Image Data Augmentation
Data augmentation in simple words means increasing the size and diver‑
sity of a dataset by creating a new set of data by applying transformations to
existing data. The main goal of data augmentation is to improve the separa‑
tion quality, and robustness of data in a generalized form so that it can be
easily used and available in a particular form.
Key features of data augmentation are as follows:
• Rotation is the rotation by a certain angle to simulate different
orientations.
• Flipping of images horizontally or vertically to create mirror images.
• Color jittering light and color of the image be adjusted and simulated.
• Data augmentation provides befits like reduced overfitting by prevent‑
ing memorizing, effectively increase the amount of training dataset and
many more.
2. Data Acquisition and Collection
The main and core point in DCAI is the acquisition of relevant and
diverse datasets.
Emerging Development and Challenges in Data‑Centric AI 31

Here, identifying and collecting high‑quality data is the main challenge

for us. These are again divided into structured data, unstructured data, and
semi‑structured data.
3. Data Preprocessing and Cleaning
Data preprocessing plays a crucial role in DCAI applications, as the
quality and suitability of the data used for training models directly impact
their performance and reliability. One of the most challenging and tedious
tasks is cleaning the data because the data does not only have a set of data
but also data with biased data which need to be sorted and separated for fur‑
ther implementation. Data preprocessing involves cleaning, transforming,
and standardizing the data to make it suitable for model training as we know
that AI models learn for meaningful patterns from the data. Its emphasis is
placed on the quality, diversity, and effective processing of data.
4. Model Training and Tuning
In DCAI, models are trained using large volumes of high‑quality data.
The iterative process of model training and hyperparameter tuning aims to
optimize the model’s performance by learning from the patterns.
5. Training data
i. Training data is a collection of data instances used to teach ML models.
Constructing high‑quality training data is critical to achieving DCAI. It is
again subdivided into five parts – Data collection (Gathering of raw data)
ii. Data labeling (Adding an extra bit of information to data samples)
iii. Data preparation (Transforming raw data to suitable form)
iv. Data reduction (Reducing data size and making it simple and
understandable)
v. Data augmentation (Increasing data diversity by creating modified sam‑
ples without collecting extra data)
6. Data Preprocessing and cleaning
DCAI approach expects the removal or correction of inaccuracies, incon‑
sistencies, and errors in the dataset to enhance data quality and convert
data into a standardized format for better model performance and reducing
biases as shown in Figure 2.2.
Nearly, every industry experiences problems with data quality, and han‑
dling them manually is incredibly time‑consuming. It becomes impossible
to guarantee the quality of datasets as they get bigger without the use of

FIGURE 2.2 Data preprocessing.

32 Data-Centric Artificial Intelligence for Multidisciplinary Applications

algorithms [9]. In an effort to address problems brought on by low‑quality

training data, recent ML systems trained on large datasets like ChatGPT
have depended heavily on labor (human feedback); nonetheless, such efforts
have not been completely successful [10]. To guarantee that ML models
are being trained with clean data, we require automated procedures and
methodical engineering principles more than before. It is crucial that these
systems are trained in a trustworthy manner as ML becomes more and more
integrated into our daily lives in the areas of healthcare, finance, and trans‑
portation. Recent studies have shown how useful DCAI is for a variety of
applications. A recent benchmark investigated various techniques to train
models under increasing noise rates in the renowned Cifar‑10 dataset [11]
for image classification with noisy labeled data. The results showed that
simple methods that adaptively alter the dataset can provide models that are
significantly more accurate than methods that attempt to account for noise
through complex modeling techniques.

2.3 EVALUATION OF DATA

The assessment of data within a DCAI framework is a vital step to ensure that the
data utilized for training and operating AI models is of high quality and relevance.
This involves a methodical evaluation of the data’s appropriateness for the intended
AI application. Key considerations encompass verifying data accuracy, distribu‑
tion, bias, and label reliability. Additionally, it involves analyzing diversity, temporal
dynamics, and potential for data augmentations. Ensuring ethical compliance and
implementing validation techniques further enhance the data’s reliability and its
impact on model performance. Through this comprehensive evaluation, organiza‑
tions can identify potential issues, ensuring the data’s integrity and the effectiveness
of AI models in real‑world scenarios.
Assessing and understanding the model quality is crucial before deployment.
Hence, there is a need to validate the data. In DCAI, two techniques are used, namely

• In‑distribution evaluation data (evaluation data refers to testing samples that

follow the same distribution as the training data.)
• Out‑distribution evaluation data (evaluation data means the testing samples
follow a distribution that differs from the training data)

2.4 DATA MAINTENANCE

Data management in DCAI involves maintaining the quality and relevance of data
used for training models. This encompasses continuous data collection, monitoring,
and cleaning. It also requires accurate labeling, data augmentation, and adapting
data to changing domains. Feedback between model performance and data quality
is important. Ethical considerations and privacy must be upheld. Documentation,
versioning, and data governance are crucial. Long‑term planning and automation
enhance sustainability. Effective data maintenance ensures AI systems provide accu‑
rate insights over time. In production scenarios, data is not created once but rather
Emerging Development and Challenges in Data‑Centric AI 33

continuously updated. Data maintenance is a significant challenge that DCAI has to

consider to ensure data reliability in a dynamic environment.

2.5 MODEL‑CENTRIC AI
In AI, a “model” is something that is referred to as a mathematical or computational
representation that understands behavior within data by capturing various patterns
and relationships. Models are the core of many AI techniques such as ML and deep
learning [12].
In model‑centric AI, the main or brute force approach is developing experimental
research to improve the ML model performance. This involves selecting the best
model architecture that will be favorable to that dataset and training process from a
wide range of possibilities as follows:

• Model Design and Architecture

• Model Interpretability
• Model transfer and adaptation
• Model optimization and training techniques
• Model ethical and responsible AI
• Model Lifecycle Management

Currently, many AI applications are based on model‑centric AI. In this, we focus on

the improvement of code or model architecture to make the use of data in a much
more efficient way. As we know, here, data is now stored in proper structure and the
proper format it makes it easy for us to get and set data which again makes it more
convenient and efficient for us to use that data. The distinction between model‑centric
AI and data‑centric AI is summarized in Table 2.1

TABLE 2.1
Comparison between Model‑Centric and Data‑Centric AI
Category Model‑Centric AI Data‑Centric AI
System Progressive upgrade of algorithm and Enhancing the data quality
Development code having constant volume and consistently while keeping the
Process fixed type of data model hyperparameters constant
Performance Excels primarily while dealing with Shows optimal performance while
substantial or extensive datasets working with a modest or compact
dataset
Robustness Susceptible to adversarial samples Higher adversarial robustness
Applicability In model‑centric AI, selecting the right Data‑centric AI often involves using
hyperparameters is critical for a variety of ML algorithms and
achieving optimal model performance models to extract insights from data
Generalization Balanced complexity of the model is Generalization starts with the
essential overlay complex model may collection of drivers and
fit the data perfectly but fails the representative data from real world
generalization to new data
34 Data-Centric Artificial Intelligence for Multidisciplinary Applications

2.6 TRANSFORMING FROM MODEL‑CENTRIC AI TO DCAI

In previous years, ML research primarily concentrated on enhancing the design of
models to achieve better performance. This involved comparing different model
architectures using established datasets, both in practical competitions and aca‑
demic studies. This approach, centered around improving models, led to significant
advancements in model architectures over time.
However, this strategy of solely refining model architectures has reached a point of
diminishing returns in terms of substantial performance improvements on well‑known
datasets. Likewise, focusing on refining complex model architectures does not
always lead to significant performance gains when dealing with real‑world datasets.
Additionally, practitioners often aim to apply ML to models. Because of this, the atten‑
tion of both researchers and practitioners is gradually shifting toward the importance
of data – the somewhat overlooked component in the development of AI systems.
Particularly, there’s a growing realization that systematic work with data is crucial
for enhancing the data used to train ML models. Data plays a pivotal role in enabling
an ML model to derive insights. Therefore, the quantity of data (such as the number
of instances) and its quality (including relevance and accuracy of labels) have a sub‑
stantial impact on the performance of AI systems. The concepts of DCAI embody
the shift in perspective. It emphasizes the significance of data as a focal point for
improving AI systems, acknowledging that the quality and quantity of data funda‑
mentally influence how well these systems perform.
DCAI and model‑centric AI differ in their core focus, the role of domain knowl‑
edge, and the understanding of data quality:

1. Focus:
In DCAI, the primary emphasis is on keeping the model architecture
constant while striving to enhance performance by improving the quality
and quantity of the data used for training.
On the other hand, model‑centric revolves around refining the model’s
design and structure to achieve better performance, with the dataset being
relatively fixed.
2. Data Work and Domain Knowledge:
DCAI demands a deep understanding of both the specific model archi‑
tecture being employed and the domain of the problem at hand. This
includes a comprehensive grasp of the underlying data. The process involves
domain‑specific data manipulation and analysis. Moreover, the develop‑
ment of techniques and tools that partially automate tasks contributes to the
creation of effective AI systems.
Model‑centric AI may not require as detailed a domain understanding,
as the primary efforts are directed toward optimizing the model’s structure.
3. Data Quality Understanding:
Modifying the foundation data raises the performance enhancements in
DCAI. Consequently, shifts in metrics used to measure the effectiveness of
ML models also reflect the impact of data adjustments. This offers a fresh
perspective on gauging data quality, approximated by changes in ML metrics.
Emerging Development and Challenges in Data‑Centric AI 35

In model‑centric AI, changes in performance metrics are more directly

attributed to adjustments in the model architecture.

2.7 CHALLENGES IN DCAI

Although very advantageous, adopting a data‑centric strategy in AI poses a num‑
ber of difficulties that businesses and practitioners must successfully overcome
to fully realize the potential of their AI models. These are the main issues with
DCAI – Modeling and Design of Data. To address these issues and lay the ground‑
work for DCAI, a multidisciplinary strategy is needed that incorporates technologi‑
cal breakthroughs, effective data management techniques, ethical considerations,
and regulatory compliance.

• Data Corruption
• Data Imbalance
• Continuous data collection and updating
• Data Fusion and Integration
• Data privacy and security
• Data bias and fairness

2.7.1 Modeling and Design of Data

One of the basic and most important things is structuring the data in the correct form
as when data we are getting in bulk form and dealing with such a huge amount of is
not so easy, and if we do not arrange it in the specific order or the correct manner,
then it will have the huge time complexity and it will be also unsecured.
Effective data modeling and design are essential for building AI systems that
deliver valuable insights and predictions. The process is iterative and may involve
experimentation with different modeling techniques and data representations to
achieve the best results for your specific problem.
Collaboration between data scientists, domain experts, and IT professionals is
often necessary to ensure a successful DCAI project.

• Data Exploration and Understanding

• Data Preprocessing
• Data Collection
• Data Splitting
• Model Selecting
• Model Training

2.7.2 Data Corruption
As we know if our data is not correct or is corrupted, then it will produce the wrong
output and lead disform of our system. So, data correctness is the most important
part, and to correct this, data on a huge scale is a big challenge because the sources
36 Data-Centric Artificial Intelligence for Multidisciplinary Applications

from which we are getting the data are not reliable, so need a strong method to make
it more successful work. Data corruption can occur at various stages of the AI life‑
cycle and can have several.
Data is a fundamental component of AI systems, and its quality directly impacts
the performance and reliability of these models. Data corruption can occur at various
stages of the AI lifecycle and can stem from several causes.

• Data collection Errors

• Data Storage and Transfer Issues
• Data Preprocessing Mistakes
• Data Labeling Errors
• Data Integration Challenges
• Data Augmentation Issues

2.7.3 Data Imbalance
In real‑world scenarios, data may be unevenly distributed across classes or catego‑
ries. It is a situation where the distribution of classes or categories in a dataset is sig‑
nificantly skewed, with some classes having much fewer instances than others. This
can lead to models that are biased toward the majority class and perform poorly on
minority classes. This imbalance can pose challenges when training and evaluating
ML models, as the models may become biased toward the majority class and perform
poorly on minority classes.

2.7.4 Continuous Data Collection and Updating

The updating of data gives a user correct information about the current situation and
continuously updates with new data to adapt to changing conditions. Also, maintain‑
ing the model’s performance as new data becomes available can be complex. To
ensure that AI models remain effective, it’s essential to establish processes for ongo‑
ing data collection, updating, and integration.
Continuous data collection and updating are integral to the success of DCAI sys‑
tems, as they help ensure that models are adaptive, accurate, and capable of providing
valuable insights in dynamic environments. Also, there is a need for reviewing and
optimizing data collection processes, which is an ongoing effort that is essential for
staying competitive and relevant in AI‑driven applications.

2.7.5 Data Fusion and Integration

In data fusion, we combine data from multiple sources, especially when dealing with
heterogeneous data and can be challenging due to differences in formats, scales, and
quality.
Also, we have seen an imbalance in that her data is heterogeneous, which may
lead to incorrect structure formation and will directly hurt the whole data format.
The main goal is to extract meaningful insights or knowledge by combining
diverse sources of information so that it is easily available.
Emerging Development and Challenges in Data‑Centric AI 37

• Sensor Data Fusion: Here, we have combined data from multiple sensors
to improve the accuracy and reliability of measurements. This is some‑
times often used in fields like robotics, autonomous vehicles, and military
applications.
• Feature‑Level Fusion: Here, we combine different features or attributes of
data from various sources. For example, combining different text data with
image data in natural language processing and computer vision applications.
• Decision‑Level Fusion– Combining the decisions or outputs of multiple
algorithms or models to make a final decision. This is common in ensemble
learning where multiple models are combined to improve.

2.7.6 Data Privacy and Security

Working with sensitive data we need to ensure that data is safe and secure.
Anonymization, encryption, and access control techniques like this should be
implemented to protect sensitive data. In this interconnected world, the security and
protection of data is one of the most important and tedious tasks, along with data
management. This involves safeguarding sensitive information from unauthorized
access, breaches, and misuse while respecting individual’s rights to control their data.

• Key Principles
• Consent: Individuals should have informed consent for the collection
and use of their data.
• Data Minimization: Collect data which need and avoid additional
access to data used on the current user login.
• Transparency: One of the main and important policies that should be
specified to the user is transparency; it should be all clear to the user and
nothing should be hidden.

2.7.7 Data Bias and Fairness

We know that AI is created on some data, if the data is biased that can lead to a
biased AI model. Developing techniques and metrics to measure and reduce bias in
AI models is one of the important challenges. The bias can lead to inaccurate and
improper predictions or reinforce existing system enforcement.

• Fairness: Fairness in ML refers to the goal of ensuring that models and algo‑
rithms treat all individuals and do not discriminate against any particular
demographic or social group. Achieving fairness is essential to prevent the
reinforcement of existing biases and to promote ethical and equitable outcomes.
• Individual Fairness: This principle aims to treat similar individuals
similarly. In other words, if any two things and these two things have
similarities in them, then the model preparation should be similar.
• Group fairness: It focuses on ensuring that predictions are fair at a
group level, particularly concerning sensitive attributes like race, gen‑
der, or ethnicity.
38 Data-Centric Artificial Intelligence for Multidisciplinary Applications

This involves avoiding disparate impact, where certain groups are disproportionately
affected by a model’s predictions.
Data bias and fairness are ongoing concerns in the field of ML, and addressing
these issues is crucial for building trustworthy and ethical AI systems that benefit
society as a whole.
Design scalable and enhanced data architecture which is the need for your AI. Use
proper database systems and data storage solutions to manage and fetch data easily
and efficiently. We can create data pipelines that self‑operate the process of clean‑
ing, transforming, and integrating data [5], which check data accuracy before setting
data into AI. We can implement a self‑operating data validation process that identifies
the inconsistencies in the incoming data. Create data recovery systems like backups
and versioning to restore clean data if it gets corrupted. Apply techniques like the
Synthetic Minority Oversampling Technique (SMOTE) to balance class distribution in
training data. Real‑time data ingestion pipelines collect real‑time data and process it,
which ensures that AI has updated data. If we can Store the data in the blockchain and
provide access to data can resolve the problem of security, as we have seen that data is
getting imbalance, we will divide the data in form of small, distributed, and structured
chunks so that the data is now in structured and distributed format; this chunks will be
the single block of this blockchain and will be accessed by the authenticated user only.

2.8 CONCLUSION
This chapter describes that model‑centric AI and DCAI have highlighted important
factors that have a big impact on the efficiency and dependability of AI models. When
developing and implementing AI systems, the problems outlined in the data‑centric
approach must be taken into account. The distinctions between model‑centric AI
and DCAI were emphasized, highlighting the significance of creating a scalable
and improved data architecture for successful AI implementation. Important tactics
include using appropriate database systems, applying data validation procedures, and
using methods like SMOTE for data balance. In addition, utilizing blockchain for
distributed, secure data storage is a viable answer to the problems with data security
and imbalance. To design AI systems that are dependable, precise, and unbiased and
eventually contribute to a more robust and fairer AI‑driven future, it is crucial to
acknowledge and effectively address these problems in DCAI.

REFERENCES
1. Zha, Daochen, Bhat, Zaid Pervaiz, Lai, Kwei‑Herng, Hu, Xia et al. (2023). Data‑centric
AI: Perspectives and challenges. doi:10.48550/arXiv.2301.04819.
2. Mingyang, Wan, Zha, Daochen, Liu, Ninghao, and Zou, Na. (2023). In‑processing
modeling techniques for machine learning fairness: A survey. ACM Transactions on
Knowledge Discovery from Data 17, 3, 1–27
3. Press, G. (2021). Andrew Ng launches a campaign for data‑centric AI. Forbes.
4. Northcutt, C., Jiang, L., and Chuang, I. L. (2021) Confident learning: Estimating uncer‑
tainty in dataset labels. Journal of Artificial Intelligence Research, 70, 1373–1411
Emerging Development and Challenges in Data‑Centric AI 39

5. Patel, Hima, Guttula, Shanmukha, Gupta, Nitin, Hans, Sandeep, Mittal, Ruhi, and
Lokesh, N. (2023). A data centric AI framework for automating exploratory data analysis
and data quality tasks. Journal of Data and Information Quality. doi:10.1145/3603709.
6. Zhang, Huaizheng, Huang, Yizheng, and Li, Yuanming. (2023). DataCI: A platform for
data‑centric AI on streaming data.
7. Verdecchia, Roberto, Cruz, Luís, Sallou, June, Lin, Michelle, Wickenden, James,
and Hotellier, Estelle. (2022). Data‑centric green AI: An exploratory empirical study.
doi:10.1109/ICT4S55073.2022.00015.
8. Khang P.H, Alex, Gujrati, Rashmi, Rani, Sita, Uygun, Hayri, and Gupta, Dr‑Shashi.
(2023). Designing workforce management systems for industry 4.0: Data‑centric and
AI‑enabled approaches. doi:10.1201/9781003357070.
9. Chiang, T. (2023). ChatGPT is a blurry JPEG of the Web, February, 2023, https://www.
newyorker.com
10. Lee, Youngjune, Kwon, Oh Joon, Lee, Haeju, Kim, Joonyoung, Lee, Kangwook and
Kim, Kee‑Eung. (2021). Augment, valuate: A data enhancement pipeline for data‑Cen‑
tric AI. In: NeurIPS Data‑Centric AI Workshop. arXiv preprint arXiv:2112.03837.
11. Redman, T. (2016). Bad data costs the U.S. $3 trillion per year. Harvard Business
Review, 22, 11–18.
12. Jarrahi, Mohammad, Memariani, Ali, and Guha, Shion. (2022). The principles of
data‑centric AI (DCAI). doi:10.48550/arXiv.2211.14611.
 3 Unleashing the Power
of Industry 4.0
A Harmonious Blend
of Data‑Centric and
Model‑Centric AI
Manivannan Karunakaran, Batri Krishnan,
D. Shanthi, J. Benadict Raja,
and B. Sakthi Karthi Durai

3.1 INTRODUCTION
Industry 4.0, often referred to as the fourth industrial revolution, has ignited a trans‑
formative wave of automation, leading to the emergence of novel industrial appli‑
cations that have the potential to reshape how humans interact with machines [1].
In response to this revolution, businesses are progressively moving away from con‑
ventional automation approaches, which heavily depend on the expertise of human
developers and the use of application programming interfaces for service platforms
[2]. The fourth industrial revolution, also known as Industry 4.0, has brought about
a transformation in engineering systems, seamlessly integrating sensing capabilities,
computational power, control, and networking into cyber‑physical objects. These
objects are interconnected through the Internet of Things (IoTs) [3].
Industry 4.0 is distinguished by its reliance on data, which plays a prominent role
across various technologies [4]. Cloud computing facilitates the effective storage,
analysis, and processing of vast amounts of data in the cloud [5]. Edge computing, on
the other hand, minimizes latency in real‑time production operations by efficiently
analyzing data near the sensors [6]. Digital twins play a crucial role in simulating var‑
ious systems’ processes in virtual environments, utilizing data from IoT sensors and
interconnected objects [7]. Artificial intelligence (AI) and machine learning (ML)
are crucial components in handling the vast data volumes of Industry 4.0 [8]. AI
systems, mainly algorithms (code), learn prototypical features from extensive data to
solve problems across different formats like text, audio, image, and video. ML, a sub‑
set of AI, enables AI systems to detect imperceptible patterns using general‑purpose
procedures, solving problems without explicit programming [9].
Data preparation involves human experts labeling and curating data for con‑
text and interpretation, known as “data annotation.” In the context of AI and ML,

40 DOI: 10.1201/9781003461500-4
Unleashing the Power of Industry 4.0 41

data‑centric AI and model‑centric AI represent two distinct approaches to improving

AI system performance [10]. Model‑centric AI focuses on iterative upgrades to the
algorithm while keeping data volumes and types constant. In contrast, data‑centric
AI emphasizes continuous improvement of data quality to achieve optimal perfor‑
mance, irrespective of the model [11]. Rather than advocating for either approach
exclusively, this chapter advocates a balanced “both/and” perspective, reconciling
data‑centric AI with model‑centric AI. It highlights the limitations of model‑cen‑
tric AI and underscores the importance of integrating data‑centric AI to overcome
these limitations [12]. Within the ML community, there are two contrasting per‑
spectives when it comes to enhancing AI system performance: model‑centric AI
and data‑centric AI. Model‑centric AI involves continuous iterations to improve the
algorithm while keeping the data volume and type constant. Conversely, data‑centric
AI emphasizes the continual enhancement of data quality to achieve superior per‑
formance in managing data noise (Figure 3.1). Both model‑centric AI and data‑cen‑
tric AI involve continuous, iterative processes, often referred to as model‑cycle and
data‑cycle approaches [13]. For the past three decades, model‑centric AI has been
the prevailing approach in both research and industry, representing over 90% of pub‑
lished AI research projects [14]. However, in recent times, there has been a notable
shift in favor of data‑centric AI. This change in perspective is driven by several
factors, such as the limited availability of large datasets and the growing demand
for customized solutions [15]. Andrew Ng, the founder of “DeepLearning.AI” and
Adjunct Professor at Stanford University, introduced the concept of data‑centric AI.
This approach demonstrated remarkable superiority over the model‑centric approach
in terms of accuracy and learning speed for detecting steel defects. Furthermore, the
data‑centric approach exhibited excellent performance in other tasks, including solar
defect detection and surface inspection [16]. Interestingly, when the audience was
asked whether to enhance the model’s accuracy by improving the code or the data,
the majority of voters in the live stream opted to improve the data, highlighting the

FIGURE 3.1 Data‑centric AI.

42 Data-Centric Artificial Intelligence for Multidisciplinary Applications

growing recognition of the importance of data‑centric AI [17]. Rather than advocat‑

ing for a rigid “either/or” perspective, this work promotes a more inclusive “both/
and” approach. While illustrating the relative effectiveness of data‑centric AI would
require rigorous experimental procedures comparing it to model‑centric AI, conduct‑
ing such a comprehensive study exceeds the current research scope. Nevertheless,
there is a discernible shift in the AI/ML community, with more researchers endors‑
ing the transition from model‑centric to data‑centric approaches. While showcas‑
ing the capabilities of data‑centric AI is crucial, it is essential to acknowledge that
isolated success stories from experimental research cannot conclusively establish its
superiority over other alternatives, including model‑centric AI. To truly demonstrate
the superiority of data‑centric AI, experimental results should be provided across
various applications, acknowledging the challenges inherent in research studies [18].

3.1.1 Contribution
This chapter significantly extends prior research [19]. We employ a comparative
­analysis methodology to contrast data‑centric AI and model‑centric AI (Table 3.1).
The analysis draws from Andrew Ng’s live stream presentation on 24 March 2021 [20].
This chapter also considers the growing number of researchers who support the tran‑
sition from model‑centric AI to data‑centric AI. The main thesis of this chapter is that
while model‑centric AI may have its limitations in terms of performance, embracing a
collaborative approach that combines both data‑centric and model‑centric methodolo‑
gies would lead to more substantial advancements in current AI technology compared
to focusing solely on improving datasets, despite the critical importance of dataset
enhancement. The key contributions of this chapter can be summarized as follows:
In Section 3.2, we present a succinct review and comprehensive discussion of the
deep learning (DL) technique and its pivotal role in propelling current AI advance‑
ments. Sections 3.4 and 3.5 establish crucial connections between current AI, cyber
security, and natural language inference (NLI). We emphasize the drawbacks of
model‑centric AI, particularly regarding its algorithmic stability and robustness.
We specifically point out examples like adversarial samples and hypothesis‑only
biases to illustrate the difficulties faced in this approach. In Section 3.6, we delve
into the motivation for adopting the data‑centric AI approach, particularly empha‑
sizing the significant impact of the IoTs’ continuous expansion, supported by the
latest relevant data. Finally, in Section 3.7, we reconcile the data‑centric AI perspec‑
tive with model‑centric AI, presenting additional arguments in favor of the “both/
and” approach over the less optimal “either/or” stance. By exploring the benefits of
combining both methodologies, we shed light on the synergistic potential of this col‑
laborative approach.

3.2 RELATED WORK

3.2.1 The Crucial Role of DL in Contemporary AI
In history, it was John McCarthy (1927–2011) who first coined the term “Artificial
Intelligence” in 1956. He provided a definition, describing AI as the “science and
Unleashing the Power of Industry 4.0 43

FIGURE 3.2 Artificial neural network with hidden layer.

engineering of creating intelligent machines, particularly intelligent computer pro‑

grams” [21]. In modern times, much of the excitement in the field of AI is centered
around DL [22], an advanced form of artificial neural network (ANN) (Figure 3.2).
In the rapid advancement of AI/ML capabilities, DL has consistently demonstrated
human‑level performance in various applications. DL has demonstrated remarkable
capabilities in various domains. It excels in tasks like object classification, outper‑
forming human performance, and defeating world‑class players in complex games
like Go and Poker. Additionally, DL has showcased its potential in medical appli‑
cations, accurately identifying cancer from X‑ray scans [23]. Furthermore, it has
proven to be effective in natural language processing tasks, facilitating text trans‑
lation across different languages. Curiously, although the theoretical groundwork
for DL was established as early as the 1940s, it wasn’t until 2012 that researchers
from the AI/ML community and related domains recognized its resemblance to the
mechanisms of the human brain. This revelation solidified DL as the technique that
most closely resembles the complex neural processes underlying human intelligence.
Their DL algorithm demonstrated significantly higher performance and accuracy
compared to previous state‑of‑the‑art algorithms, solidifying DL’s position as a
groundbreaking technique in the field of AI. This victory in the ImageNet contest
highlighted the transformative potential of DL and its ability to outperform tradi‑
tional approaches in complex tasks, further fueling the excitement surrounding this
powerful AI methodology.
In a typical ANN, learning takes place by adjusting the “weights,” represented as
w, that amplify or attenuate signals denoted by x, which are transmitted through con‑
nections between nodes in the ANN. The ANN contains several hidden layers, resem‑
bling processing centers akin to those in the human brain and employs backpropagation
of gradients. In each layer of ANNs, the total input z is calculated as the weighted
sum of the outputs from the previous layer. Subsequently, a nonlinear function f(.) is
applied to z, generating the output of the unit. Commonly used nonlinear functions in
ANNs include the rectified linear unit (ReLU) f(z) = max(0, z), as well as traditional
sigmoids such as the hyperbolic tangent f(z) = (exp(z) − exp(−z))/(exp(z) + exp(−z))
44 Data-Centric Artificial Intelligence for Multidisciplinary Applications

and the logistic function f(z) = 1/(1 + exp(−z)) [3]. These activation functions intro‑
duce nonlinearity to the neural network, enabling it to model complex relationships
and learn from data more effectively.
In contrast to traditional ANNs, DL networks acquire data representations at vari‑
ous levels of abstraction. The achievement of modeling complex relationships in neu‑
ral networks is made possible by employing a significantly higher number of layers.
These layers consist of simple, nonlinear modules known as neurons, which play a
pivotal role in transforming the internal representation of specific input data aspects
from one layer to a higher‑level internal representation. As data passes through mul‑
tiple layers, the neural network learns to abstract and represent increasingly com‑
plex features, allowing it to capture intricate patterns and relationships in the data.
This hierarchical approach to representation learning enables deep neural networks
to tackle challenging tasks and achieve remarkable performance in various appli‑
cations. The backpropagation algorithm fine‑tunes the neural network’s weights by
working back down through the layers, adjusting each weight proportionally based
on its contribution to the overall error. DL networks are characterized by their sig‑
nificant depth, typically ranging from 5 to 20 layers, which is what earned them the
name “deep” networks. However, in modern commercial applications, neural net‑
work models frequently employ over 100 layers, emphasizing their ability to scale
too much larger depths. This substantial depth enables DL models to learn hierarchi‑
cal representations and abstract complex features from raw input data, allowing them
to handle intricate tasks effectively. In contrast to classical ML techniques, which
rely on human experts to carefully engineer features and extract relevant aspects of
input data, DL models possess the ability to autonomously learn these representa‑
tions directly from the data. This process, known as feature learning or represen‑
tation learning, allows DL models to operate without explicit human intervention
in the feature extraction step. This characteristic is particularly advantageous as it
reduces the burden of manual feature engineering and enables DL models to adapt
to a wide range of tasks and datasets, making them highly flexible and powerful
tools in various applications. DL algorithms implicitly learn features from data using
general‑purpose procedures. The exceptional capability of DL has drawn compari‑
sons to the problem‑solving approach of the human brain. Additionally, a significant
finding in the field is that ANNs, with a non‑polynomial activation function, pos‑
sess the ability to approximate any continuous function with arbitrary accuracy. This
finding, known as the universal approximation theorem, mathematically establishes
the equivalence of ANNs to universal computers. In essence, this theorem confirms
that DL models can represent and approximate a wide range of complex functions,
making them highly versatile and powerful tools for solving diverse problems. This
unique combination of human brain‑like problem‑solving and mathematical prowess
has contributed to the widespread success and adoption of DL in numerous domains
and applications.
The evolution of DL into its current state and the fulfillment of many early aspi‑
rations of AI researchers can be attributed to several significant factors. A critical
aspect is DL’s ability to excel when presented with abundant and diverse data from
various sources in the Global Data sphere. This extensive data sphere encompasses
textual, visual, and acoustic data and comprises a wide range of sources, including
Unleashing the Power of Industry 4.0 45

FIGURE 3.3 Projected annual data creation, consumption, and storage according to IDC.

cloud data centers, enterprise infrastructure, endpoints such as PCs, smartphones,

sensors, social media platforms, wearable devices, and many more. The availability
of such large volumes of data from diverse origins has been instrumental in fueling
the advancement and success of DL, enabling it to achieve exceptional performance
across a multitude of applications and domains. DL algorithms require data to train
models capable of identifying statistical similarities between current and previously
observed instances to uncover hidden patterns through unsupervised learning and
make future predictions for unseen instances through supervised learning. The per‑
formance and accuracy of the underlying algorithm increase with more granular,
voluminous, and diverse data. As per the International Data Corporation (IDC), a
renowned market intelligence company, the Global Data sphere is anticipated to
experience an almost exponential growth (Figure 3.3). Notably, the volume of data
created and replicated in 2010 was relatively modest, not surpassing 1 zettabyte (ZB).
However, by the year 2020, this volume had expanded significantly, reaching 64.2
ZB, and it is projected to exceed 180 ZB by 2025 (1 ZB = 10^12 gigabytes). These
staggering figures indicate the rapid expansion of data generation and replication
on a global scale, highlighting the substantial growth potential and the immense
data‑driven opportunities that lie ahead for DL and other AI techniques.

3.3 GAP ANALYSIS

The book chapter “Driving Compact and Robust Industry 4.0 Solutions: The
Synergy between Data‑Centric and Model‑Centric AI” delves into the integra‑
tion of data‑centric and model‑centric approaches in developing Industry 4.0 solu‑
tions. The chapter emphasizes the significance of both methodologies and proposes
46 Data-Centric Artificial Intelligence for Multidisciplinary Applications

a complementary interplay to optimize AI‑driven solutions within the context of

Industry 4.0. While the chapter provides valuable insights, several gaps in its con‑
tent could be addressed to enhance its comprehensiveness and practicality. First, the
chapter could benefit from more in‑depth case studies that illustrate the successful
application of data‑centric and model‑centric AI in real‑world Industry 4.0 scenarios.
These case studies would provide practical examples of how these approaches have
been deployed, the challenges faced, and the solutions derived from their integration.
Additionally, a comprehensive comparative analysis of the strengths and limitations of
data‑centric and model‑centric AI would offer valuable insights into their individual
contributions and the potential benefits of combining them. Another area that requires
further exploration is the consideration of ethical aspects and data biases in AI appli‑
cations. While the chapter briefly mentions the importance of addressing data biases,
a more extensive discussion on techniques to mitigate biases and ensure fairness in
AI models would be beneficial. Moreover, the challenges of real‑time data processing
in Industry 4.0 environments are crucial to address, as agile decision‑making relies
on timely feedback. Understanding the computational and architectural requirements
for real‑time data processing in both data‑centric and model‑centric AI models is
essential for their practical implementation. The book chapter highlights the impor‑
tance of integrating data‑centric and model‑centric AI approaches for robust Industry
4.0 solutions. However, addressing the identified gaps, such as the need for in‑depth
case studies, a comparative analysis, ethical considerations, and real‑time data pro‑
cessing challenges would enhance the chapter’s insights. By bridging these gaps, the
chapter can offer valuable guidance to researchers, practitioners, and industry leaders
seeking to harness the full potential of AI in Industry 4.0.

3.4 CHALLENGES FACED BY MODEL‑CENTRIC AI

Model‑centric AI primarily focuses on optimizing model architectures (algorithm/
code) and their hyperparameters, while keeping the collected data fixed throughout
the AI system’s development. However, this approach faces significant limitations,
including Model‑centric AI performs well in businesses and industries with gener‑
alized solutions, where a single AI system can satisfy the majority of users, leaving
outliers as negligible. Companies like Google and Facebook with vast standardized
data in the advertising industry benefit from such solutions. However, industries like
manufacturing, agriculture, and healthcare require tailored solutions, which cannot
be provided by standardized models. Customized approaches, encompassing com‑
prehensive and consistently labeled data, are essential for these sectors. Model‑centric
AI, especially DL networks, is susceptible to adversarial samples – instances with
small, deliberate perturbations that cause false predictions. This phenomenon raises
concerns about the stability and robustness of DL networks, potentially leading to
security issues in various applications. For instance, in autonomous cars, adversar‑
ies can manipulate DL models, leading to dangerous situations. Healthcare applica‑
tions relying on DL systems may also be compromised, affecting patient outcomes
and costs. Model‑centric AI algorithms may achieve high performance by exploit‑
ing annotation artifacts – context‑free associations created during the dataset anno‑
tation process. In NLI, human annotators introduce cognitive heuristics, resulting
Unleashing the Power of Industry 4.0 47

in datasets with biases. While the model performs well on such datasets, it lacks
generalization to new data without similar biases, demonstrating its low generaliza‑
tion capacity.
Prior to the widespread recognition of “data‑centric AI,” AI researchers and ML
practitioners devoted significant efforts to curating datasets for training ML models.
While some data instances were discovered to be invalid due to mislabeling, ambi‑
guity, or irrelevance, their influence on model performance was often deemed insig‑
nificant. However, businesses and industries have now come to realize the crucial
significance of prioritizing high‑quality datasets throughout the entire development
process of AI systems.
The foremost challenge lies in the scarcity of sufficiently large and diverse datasets.
Unlike internet companies, manufacturing industries often have limited data, with
training datasets containing only thousands of relevant data points. Consequently,
ML models built on such limited data struggle to perform effectively when com‑
pared to models trained on massive datasets. In industries like manufacturing, where
a variety of products are produced, a one‑size‑fits‑all AI system for fault detection
may not be sufficient. Each product demands its own uniquely trained ML system to
ensure effective performance and accurate fault detection.

3.5 CHARACTERISTICS OF DATA‑CENTRIC AI

Data‑centric AI involves continuous evaluation of the AI model along with data
updates (Table 3.1). Unlike model‑centric AI, where the model is trained on a dataset
only once during production, data‑centric AI assumes successive improvements in
data. This is especially beneficial for businesses and industries with limited data
points, such as manufacturing, agriculture, and healthcare. The data‑centric approach
allows the model to encounter novel data instances during its lifespan, leading to
more frequent model assessments and the ability to recognize and adapt to distri‑
butional data drifts. According to Statista and Transforma Insights, the number of
IoT‑connected devices worldwide is anticipated to undergo significant growth, with
projections showing a tripling from 9.7 billion in 2020 to over 29 billion in 2030
and an increase from 7.6 billion in 2019 to 24.1 billion in 2030, respectively. This
substantial growth is attributed to the escalating impact of digital transformation,
further amplified by the effects of the COVID‑19 pandemic. The increasing adoption
of IoT technologies is driving this exponential growth, transforming industries across
the globe and shaping the future of AI and data‑centric AI applications.
The implementation of a data‑centric approach involves a series of interconnected
steps. It demands the integration of diverse expertise from different disciplines to
empower AI systems and achieve optimal performance. In the following sections, we
delineate several practical rules and criteria that play a pivotal role in successfully
embracing a data‑centric approach, ultimately leading to the development of more
effective and efficient AI systems. One essential aspect of achieving a data‑centric
approach is curating datasets to ensure both the sufficiency and representativeness
of data inputs. To achieve this, the devised model should access a wide range of
task‑relevant data inputs, including relevant information and noise present in real‑life
situations. Similar to the human brain’s selective attention, the model should focus
48 Data-Centric Artificial Intelligence for Multidisciplinary Applications

on goal‑relevant aspects while filtering out distracting and noisy information. The
success of AI models heavily relies on the quality of the data used for their train‑
ing. In the data‑centric AI paradigm, the focus shifts from merely fine‑tuning the
model’s architecture and hyperparameters to prioritizing the acquisition of sufficient
and representative data inputs. This chapter aims to explore the significance of ade‑
quate and representative data inputs for data‑centric AI and how they contribute to
achieving peak performance and robustness in AI systems. Data is the lifeblood of
AI systems, serving as the foundation upon which models learn to make decisions
and predictions. In data‑centric AI, the emphasis is on obtaining datasets that not
only contain a large volume of data but also encompass a diverse range of examples
that reflect real‑world scenarios. Adequate data inputs are crucial to ensure that AI
models have enough information to learn complex patterns and relationships within
the data, making them capable of solving the specific task at hand. The sufficiency
of data inputs plays a pivotal role in the effectiveness of data‑centric AI. Insufficient
data can lead to underfitting, where the model fails to capture the underlying patterns
in the data, resulting in poor performance. To ensure data sufficiency, researchers
and practitioners need to carefully curate datasets that encompass a broad spectrum
of instances relevant to the task. This includes not only positive examples but also
negative and ambiguous instances, providing the model with a comprehensive under‑
standing of the problem space. In addition to sufficiency, the data inputs must also
be representative of the real‑world scenarios that the AI system will encounter. A
dataset that is biased or lacks diversity can lead to a model that performs well on the
training data but fails to generalize to unseen data from different distributions. To
address this, researchers must be vigilant in avoiding bias during data collection and
ensure that the dataset accurately reflects the target population.
Curating datasets with adequate and representative data inputs can be challeng‑
ing, especially in domains where data is scarce or subject to privacy constraints.
However, there are several strategies to overcome these challenges. Collaborating
with domain experts and stakeholders can help in identifying critical data attributes
and real‑world use cases. Data augmentation techniques can also be employed to
increase the diversity of the dataset, enabling the model to learn from a wider range
of instances. Adequate and representative data inputs are fundamental to the success
of data‑centric AI. By prioritizing the quality of data during the AI model develop‑
ment process, researchers and practitioners can create more robust and effective AI
systems capable of generalizing to real‑world scenarios. As the AI field continues to
advance, the importance of data‑centric AI will only grow, making it imperative for
data scientists and AI engineers to focus on obtaining high‑quality data inputs for
their models.
Another key step in implementing a data‑centric approach is ensuring high‑quality
data during data preparation. Research teams must be cautious of potential biases
introduced during data labeling. To address this, textual descriptions can be incor‑
porated as an intermediate step between data inputs and label assignments. These
descriptions, consisting of 3–10‑word sentences, provide contextual information
reflecting human perspectives. Although this approach may extend data creation time,
it proves beneficial for AI engineers as it ensures the collected data vividly captures
the essential concepts required for effective learning by AI systems. Consequently,
Unleashing the Power of Industry 4.0 49

AI systems can efficiently learn from smaller datasets, a common scenario in vari‑
ous industries. The third step entails the continuous engagement of both AI‑ and
business‑domain experts. Domain experts should take charge of data engineering
as they possess in‑depth knowledge of specific business use cases, enabling them to
provide domain‑specific representations of the real world. Their involvement in the
evaluation process can enhance it significantly by designing domain‑sensitive tests
for the AI model, making AI more applicable and accessible across various indus‑
tries. The fourth step involves the implementation of MLOps (Machine Learning
Operations) platforms. By leveraging these platforms, research teams can reduce the
time and effort spent on software development, leading to a decrease in the main‑
tenance cost of AI applications. MLOps platforms offer essential software scaffold‑
ing for the production of AI systems, considerably reducing the time from proof of
concept to production, transforming the timeline from years to mere weeks. These
platforms encompass a range of MLOp tools that cater to both data‑centric and
model‑centric AI, including data labeling, data cleaning, model storage, continuous
integration, training, and deployment tools. Their adoption streamlines the develop‑
ment and deployment process, fostering efficient and scalable AI applications.

3.7 RECONCILING DATA‑CENTRIC AI WITH MODEL‑CENTRIC AI

At first glance, the instinct may be to concentrate on refining the model (algorithm/
code), rather than enhancing the data to improve the performance and robustness of
an AI system. Nevertheless, taking into account the constraints of model‑centric AI
and the challenges encountered by DL‑based models, as discussed in Section 3.5
(or summarized in Table 3.1), the significance of data‑centric approaches becomes

TABLE 3.1
Distinguishing Traits of Model‑Centric AI and Data‑Centric AI
Category Model‑Centric AI Data‑Centric AI
• System development • The iterative enhancement of a • Consistent enhancement in the
lifecycle model (algorithm/code) using a data quality with unchanging
fixed volume and type of data model hyperparameters
• Performance • Demonstrates high performance • Demonstrates strong performance
primarily with extensive datasets even with smaller datasets
• Robustness • Vulnerable to adversarial • Exhibits higher resilience against
samples adversarial samples
• Applicability • Suitable for evaluating • Especially well‑suited for
algorithmic solutions in real‑world scenarios and
applications with specific and applications
limited tasks
• Generalization • Limited ability to generalize • Likely to achieve good
across datasets due to a lack of generalization across various
contextual understanding datasets beyond the ones used for
testing
50 Data-Centric Artificial Intelligence for Multidisciplinary Applications

apparent. We advocate for a balanced approach, recognizing that both model‑centric

and data‑centric methods are interconnected aspects that influence each other.
Embracing this interconnectedness allows for a more comprehensive and effective
AI development strategy, fostering advancements in performance and robustness
while capitalizing on the strengths of both approaches.
While the idea of transitioning to data‑centric AI due to model‑centric limitations
is appealing, creating new datasets can be costly and may not guarantee freedom from
new artifacts or biases. Human annotators might still introduce subtle cognitive strat‑
egies, affecting the model’s generalization capability. Therefore, a better approach is
to simultaneously improve both the model and data, recognizing their interdepen‑
dence and treating them as complementary aspects. Effective problem‑solving in AI
systems necessitates a holistic consideration of both the “how‑to” (model) and the
“what‑is” (data) aspects. Just as our natural problem‑solving approach involves both
action and understanding properties, AI development also requires a balanced focus
on algorithm design and hyperparameter optimization to devise effective models.
Simultaneously, the utilization of a consistent dataset is essential for accurate model
comparison and comprehensive performance evaluation. By integrating these com‑
ponents, AI developers can create robust and high‑performing AI systems, leveraging
the interplay between algorithms and datasets to achieve optimal problem‑solving
capabilities.
Conversely, adopting a model‑centric approach offers valuable opportunities to
accumulate experience in real‑world problem‑solving and computational solutions.
By taking action and gaining insights, developers can develop a deeper understand‑
ing of the environment, analogous to how intelligent agents learn through reinforce‑
ment learning, a celebrated ML technique [6,8]. This process of iterative learning
and continuous improvement allows AI systems to refine their problem‑solving
capabilities over time, adapting to complex and dynamic environments effectively.
The model‑centric approach serves as a stepping stone for honing AI systems’ per‑
formance and enabling them to achieve increasingly sophisticated problem‑solving
skills. Indeed, the limitations of models do not necessarily equate to the limitation
of modeling itself. Aristotle’s first‑principles thinking underscores the importance of
understanding fundamental aspects to discover effective solutions. In the context of
the NLI task, models might learn from annotation artifacts, resulting in high perfor‑
mance on biased datasets but faltering on unbiased ones. This observation highlights
the significance of being mindful of the underlying biases and limitations of the data
when designing AI models. It also emphasizes the need for more comprehensive
and unbiased datasets to foster the development of robust and unbiased AI systems
capable of generalizing effectively to diverse real‑world scenarios.
Absolutely, acknowledging the limitations of models does not imply that the
model‑centric approach is inherently restricted, especially in the context of NLI. A
deeper understanding of how artifacts emerge from disregarding premise‑hypothesis
relationships can inform the design of models to mitigate such learning issues. For
instance, Belinkov et al. proposed an innovative approach by inputting both the
hypothesis and the entailment label to predict the premise. This diverges from the con‑
ventional NLI models that predict an entailment label based on a premise‑hypothesis
pair. By adopting novel strategies that capture and exploit crucial relationships,
Unleashing the Power of Industry 4.0 51

model‑centric AI can overcome certain limitations, leading to more accurate and

robust performance in NLI tasks and other applications. This showcases the dynamic
nature of the model‑centric approach, which can evolve and adapt to address specific
challenges and achieve improved results in various AI tasks. To enhance the robust‑
ness of DL/ML models and foster a harmonious interplay between data‑centric and
model‑centric AI, aligning algorithms with nature’s intuitive and powerful mecha‑
nisms, such as the learning and problem‑solving mechanisms observed in the human
brain, is essential. One such mechanism is common sense, which pertains to the abil‑
ity of humans (and animals) to acquire world models by accumulating vast contex‑
tual knowledge about the functioning of the world and the causal relationships that
govern events. This knowledge is then employed to predict future outcomes.
By incorporating common‑sense reasoning into AI models, we can infuse them
with a more human‑like understanding of the world, allowing them to adapt to vari‑
ous situations with greater accuracy and robustness. This interdisciplinary approach,
drawing inspiration from the complexities of the human brain, holds great potential for
advancing AI systems and bridging the gap between data‑centric and model‑centric
methodologies. Emulating the intuitive mechanisms found in nature can lead to
more sophisticated and comprehensive AI systems capable of achieving greater
problem‑solving capabilities and generalizing effectively across diverse scenarios.
Indeed, current AI and ML systems often lack the innate ability of common‑sense
reasoning that humans possess. For instance, while a human driver can intuitively
predict the negative outcome of driving too fast based on their understanding of
physics and prior experiences, an AI system in an autonomous vehicle would require
thousands of reinforcement learning trials to acquire similar knowledge. This dis‑
crepancy highlights the disparity between human intelligence, which relies on accu‑
mulated knowledge and affordances provided by the brain during problem‑solving,
and AI systems, which typically require extensive data and learning iterations to
reach comparable conclusions. Humans can adapt and navigate new situations by
leveraging their interactions with the environment and their inherent cognitive abili‑
ties. In contrast, AI systems often require vast amounts of labeled data and complex
learning algorithms to handle even seemingly simple scenarios. Bridging this gap
between human intelligence and AI remains a significant challenge in the field of AI
research. Integrating common‑sense reasoning and leveraging human‑like cognitive
affordances could potentially lead to the development of more robust and versatile
AI systems capable of handling new and unforeseen situations with greater efficiency
and accuracy.
The use of adversarial samples in cyber attacks, where data can be freely submit‑
ted into a running DL/ML algorithm, underscores the need for a complementary
approach involving both data‑centric and model‑centric AI. Initially, false predictions
in DL networks with perturbed dataset samples were attributed to extreme nonlin‑
earity, insufficient model averaging, and inadequate regularization in purely super‑
vised learning. This highlights the importance of incorporating both data‑centric
and model‑centric methodologies to address vulnerabilities and enhance AI system
robustness. Subsequently, a deeper understanding revealed that adversarial samples
actually arise from the inherent linear characteristics of neural networks in gen‑
eral. Adversarial samples are not random artifacts arising from typical variability
52 Data-Centric Artificial Intelligence for Multidisciplinary Applications

during propagation learning runs, nor are they caused by overfitting or incomplete
model training. Instead, they demonstrate resilience to random noise and can trans‑
fer between neural network models, even when these models differ in the number of
layers, hyperparameters, and training data. This remarkable characteristic empha‑
sizes the robustness challenges faced by AI systems and underscores the impor‑
tance of considering both data‑centric and model‑centric approaches to enhance
their performance and security. This suggests that the robustness of a deep neural
network model based on backpropagation is not solely determined by the datasets
used for training. Instead, it is influenced by the structural connection between the
network and the data distribution. Therefore, a combination of both data‑centric and
model‑centric approaches is essential when striving to enhance the network’s robust‑
ness. By addressing both aspects, AI researchers and practitioners can create more
resilient and secure neural network models.

3.8 CONCLUSIONS
The AI/ML community is witnessing a surge of interest in data‑centric AI, poten‑
tially heralding a paradigm shift in AI/ML model development. Our thorough
analysis explored the advantages and disadvantages of both data‑centric and
model‑centric approaches, culminating in a balanced perspective that harmonizes
the two approaches. While we concur with fellow researchers on the importance of
data‑centric AI, we firmly assert that this shift should not diminish the significance
of model‑centric AI. Embracing the data‑centric approach is a crucial advancement
in AI capabilities, as it places a strong emphasis on high‑quality datasets, enabling
models to better comprehend and predict real‑world scenarios.
However, it is equally vital not to overlook the significance of model‑centric
AI, which concentrates on refining and optimizing the algorithms and hyperpa‑
rameters that form the backbone of AI models. This approach has proven highly
valuable in achieving superior performance and efficiency across various appli‑
cations. Achieving a balance and recognizing the complementary nature of both
data‑centric and model‑centric AI is the key. By amalgamating the strengths of
both approaches, we can forge more resilient and potent AI systems that excel in
diverse situations. The context of Industry 4.0, promising revolutionary automa‑
tion and IoT‑driven interaction among cyber‑physical objects, accentuates the criti‑
cal interplay between data‑centric and model‑centric AI. The triumph of Industry
4.0 technologies hinges on effective data utilization and sophisticated AI models.
Merging both approaches empowers us to fully leverage AI’s potential in Industry
4.0 applications, leading to heightened efficiency and innovation across multiple
domains.
In conclusion, embracing the data‑centric AI approach represents significant
progress, but it must not overshadow the importance of model‑centric AI. Instead,
recognizing the value of both approaches and synergistically incorporating them can
usher in groundbreaking AI technologies, particularly in the context of Industry 4.0’s
ambitious goals. By striking a harmonious balance, we pave the way for a new era of
AI development, where data‑centric and model‑centric AI unite to reshape the fron‑
tiers of AI, transcending limitations and unlocking untapped potential.
Unleashing the Power of Industry 4.0 53

3.9 FUTURE WORK

In the realm of data‑centric AI and model‑centric AI, future endeavors should focus
on maximizing the potential of AI technologies through various avenues. First,
advancements in data‑centric AI demand refinement of data preparation techniques
to ensure datasets are sufficient and representative, along with strategies to identify
and mitigate biases. Exploring unsupervised and self‑supervised learning methods
can further enrich AI systems with common‑sense knowledge. Second, integrating
human domain expertise into data engineering and model development will lead to
tailored and effective AI solutions.
In the realm of model‑centric AI, continuous improvement of DL architectures,
optimization algorithms, and regularization techniques will enhance model perfor‑
mance, robustness, and interpretability. MLOps and deployment platforms can sig‑
nificantly streamline the development and deployment of AI applications, making
them more efficient and agile. Exploring hybrid approaches that combine data‑driven
and model‑driven optimization can lead to more robust and versatile AI systems.
Additionally, addressing ethical and social implications, such as developing guide‑
lines for responsible AI deployment, fairness, transparency, and privacy protection,
is paramount. Establishing standardized benchmarks and metrics for fair AI model
comparisons, as well as real‑world deployment and validation in collaboration with
industries, will provide valuable insights for further improvements in AI technolo‑
gies. By embracing these directions, the AI community can unlock the full potential
of data‑centric and model‑centric AI paradigms and drive innovation in the field.

REFERENCES
[1] Bender, E.M.; Gebru, T.; McMillan‑Major, A.; Shmitchell, S. On the dangers of sto‑
chastic parrots: Can language models be too big? In: Proceedings of the 2021 ACM
Conference on Fairness, Accountability, and Transparency, Toronto, ON, Canada,
3–10 March 2021
[2] Bhatt, S. The Big Fight: RPA Versus Traditional Automation. 2018. https://www.
botreetechnologies.com/blog/the‑big‑fight‑robotic‑process‑automation‑vs‑traditional‑
automation (accessed on 3 January 2023).
[3] Boubin, J.; Banerjee, A.; Yun, J.; Qi, H.; Fang, Y.; Chang, S.; Srinivasan, K.; Ramnath,
R.; Arora, A. PROWESS: An Open Testbed for Programmable Wireless Edge Systems.
Association for Computing Machinery, New York, NY, USA, 2022
[4] Chen, T.; Moreau, T.; Jiang, Z.; Zheng, L.; Yan, E.; Shen, H.; Cowan, M.; Wang, L.; Hu, Y.;
Ceze, L. et al. {TVM}: An automated {End‑to‑End} optimizing compiler for deep learn‑
ing. In: Proceedings of the 13th USENIX Symposium on Operating Systems Design and
Implementation (OSDI 18), Carlsbad, CA, USA, 8–10 October 2018, pp. 578–594.
[5] Fujita, H. AI‑based computer‑aided diagnosis (AI‑CAD): The latest review to read first.
Radiological Physics and Technology 2020, 13, 6–19.
[6] Hack, U. What Is the Real Story Behind the Explosive Growth of Data? 2021. https://
www.red‑gate.com/blog/database‑development/whats‑the‑real‑story‑behind‑the‑e
xplosive‑growth‑of‑data (accessed on 3 January 2023).
[7] Hamid, O.H.; Braun, J. Reinforcement learning and attractor neural network mod‑
els of associative learning. In: Computational Intelligence: Proceedings of the 9th
International Joint Conference, IJCCI 2017, Funchal–Madeira, Portugal, 1–3 November
2017, Revised Selected Papers, Springer, New York, NY, USA, 2019, pp. 327–349.
54 Data-Centric Artificial Intelligence for Multidisciplinary Applications

[8] Hamid, O.H. From model‑centric to data‑centric AI: A paradigm shift or rather a com‑
plementary approach? In: Proceedings of the 2022 8th International Conference on
Information Technology Trends (ITT), Dubai, United Arab Emirates, 25–26 May 2022,
IEEE, Piscataway, NJ, USA, 2022, pp. 196–199
[9] Tian, Q.; Yang, Y.; Lin, C.; Li, Q.; Shen, C. Improving Adversarial Robustness with
Data‑Centric Learning. 2022. https://alisec‑competition.oss‑cn‑shanghai.aliyuncs.
com/competition_papers/20211201/rank5.pdf
[10] Jiang, Y.; Zhu, Y.; Lan, C.; Yi, B.; Cui, Y.; Guo, C. A unified architecture for accelerat‑
ing distributed {DNN} training in heterogeneous {GPU/CPU} clusters. In: Proceedings
of the 14th USENIX Symposium on Operating Systems Design and Implementation
(OSDI 20), 4–6 November 2020, pp. 463–479.
[11] Kotsiopoulos, T.; Sarigiannidis, P.; Ioannidis, D.; Tzovaras, D. Machine learning and
deep learning in smart manufacturing: The smart grid paradigm. Computer Science
Review 2021, 40, 100341.
[12] Krishnan, S.; Wang, J.; Wu, E.; Franklin, M.J.; Goldberg, K. Activeclean: Interactive
data cleaning for statistical modeling. Proceedings of the VLDB Endowment 2016, 9,
948–959.
[13] Mazumder, M.; Banbury, C.; Yao, X.; Karlaš, B.; Rojas, W.G.; Diamos, S.; Diamos, G.;
He, L.; Kiela, D.; Jurado, D. et al. DataPerf: Benchmarks for data‑centric AI devel‑
opment. 2022. Dataperf: Benchmarks for data‑centric ai development. arXiv preprint
arXiv:2207.10062.
[14] Miranda, L.J. Towards Data‑Centric Machine Learning: A Short Review. https://ljvmi‑
randa921.github.io/notebook/2021/07/30/data‑centric‑ml
[15] Molnar, C. Interpretable Machine Learning. 2022. https://christophm.github.io/inter‑
pretable‑ml‑book (accessed on 3 January 2023).
[16] Pareek, P.; Shankar, P.; Pathak, M.P.; Sakariya, M.N. Predicting music popularity
using machine learning algorithm and music metrics available in Spotify. Center for
Development Economics Studies 2022, 9, 10–19.
[17] Reinsel, D.; Rydning, J.; Gantz, J.F. Worldwide Global Datasphere Forecast, 2021–2025:
The World Keeps Creating More Data‑Now, What Do We Do With it All? 2021. https://
www.marketresearch.com/IDC‑v2477/Worldwide‑Global‑DataSphere‑Forecast‑
Keeps‑14315439/
[18] Renggli, C.; Karlaš, B.; Ding, B.; Liu, F.; Schawinski, K.; Wu, W.; Zhang, C. Continuous
integration of machine learning models with ease. ml/ci: Towards a rigorous yet practi‑
cal treatment. Proceedings of Machine Learning and Systems 2019, 1, 322–333.
[19] Russell, B.C.; Torralba, A.; Murphy, K.P.; Freeman, W.T. LabelMe: A database and
web‑based tool for image annotation. International Journal of Computer Vision 2008,
77, 157–173.
[20] Schlegl, T.; Stino, H.; Niederleithner, M.; Pollreisz, A.; Schmidt‑Erfurth, U.; Drexler,
W.; Leitgeb, R.A.; Schmoll, T. Data‑centric AI approach to improve optic nerve head
segmentation and localization in OCT en face images. arXiv preprint arXiv:2208.03868.
[21] van Moorselaar, D.; Slagter, H.A. Inhibition in selective attention. Annals of the New
York Academy of Sciences 2020, 1464, 204–221.
[22] Vartak, M.; Subramanyam, H.; Lee, W.E.; Viswanathan, S.; Husnoo, S.; Madden, S.;
Zaharia, M. ModelDB: A system for machine learning model management. In:
Proceedings of the Workshop on Human‑In‑the‑Loop Data Analytics, San Francisco,
CA, USA, 26 June–1 July 2016, pp. 1–3.
[23] Zhang, H.; Li, Y.; Huang, Y.; Wen, Y.; Yin, J.; Guan, K. MLmodelCI: An automatic
cloud platform for efficient MLaaS. In: Proceedings of the 28th ACM International
Conference on Multimedia, Seattle, WA, USA, 12–16 October 2020, pp. 4453–4456.
 4 Data‑Centric AI
Approaches for
Machine Translation
Chandrakant D. Kokane, Pranav Khandagale,
Mehul Ligade, Shreeyash Garde, and Vilas Deotare

4.1 INTRODUCTION
Machine translation, a fundamental component of interlingual communication in
our increasingly globalized world, has witnessed significant advancements in recent
years.1 As the demand for accurate and efficient translation grows, the integration of
machine learning techniques has emerged as a transformative force in enhancing the
quality and efficacy of machine translation systems.
The automated translation of text or speech from one language into another is
referred to as machine translation. It makes cross‑cultural communication easier
in a number of different professions, such as business, diplomacy, and academia.
Historically, rule‑based and statistical approaches have been utilized to address the
challenges of translation. These approaches could not adequately handle the intrinsic
language complexity, resulting in fewer accurate translations and poor translation
quality.
With the advent of machine learning, a revolutionary approach that focuses on the
development of algorithms capable of learning and improving from data, a paradigm
shift has occurred in the field of machine translation. Machine learning algorithms,
powered by neural networks, have demonstrated exceptional capabilities in capturing
the complex patterns and linguistic nuances present in diverse language pairs.
By leveraging large‑scale parallel corpora and sophisticated neural network archi‑
tectures, machine learning techniques enable machine translation systems to com‑
prehend and generate translations that exhibit higher fidelity to the original meaning
and context. This paradigm shift toward data‑driven, machine learning‑based trans‑
lation systems, commonly known as neural machine translation, has led to remark‑
able advancements and a significant improvement in translation quality.
The incorporation of machine learning in machine translation not only addresses
the limitations of traditional approaches but also opens up new avenues for explo‑
ration and innovation.1 The ability of machine learning models to learn from vast
amounts of training data and adapt to different language pairs has fueled the progress
in achieving state‑of‑the‑art translation performance.
The key contributions of the book chapter aim to explore the pivotal role of
machine learning techniques in advancing machine translation. By delving into

DOI: 10.1201/9781003461500-5 55
56 Data-Centric Artificial Intelligence for Multidisciplinary Applications

the fundamental concepts and cutting‑edge developments in the field, we will

examine the transformative impact of machine learning algorithms on translation
accuracy, fluency, and efficiency. Furthermore, we will discuss the challenges and
future directions in machine translation, paving the way for a deeper understand‑
ing and broader adoption of machine learning techniques in this multidisciplinary
domain.
Through this exploration, we strive to contribute to the growing body of knowl‑
edge surrounding machine translation and to inspire further research and innovation
in this crucial area of language technology.

4.2 MACHINE TRANSLATION SYSTEMS OVERVIEW

Over time, machine translation systems have developed, moving from statistical and
rule‑based methods to the more modern paradigm of neural machine translation.2
This section gives a general overview of these many kinds of machine translation
systems, emphasizing their unique qualities and the necessity of using machine
learning methods and the overview of the machine translation systems given in
Figure 4.1.
Rule‑based machine translation (RBMT) systems translate texts based on a preset
set of linguistic rules and patterns.2 These systems use multilingual dictionaries and
lexicons together with manual encoding of grammatical and syntactic rules. Despite
its early popularity, RBMT systems have limitations since developing rules is labor
intensive and it is difficult to account for all linguistic differences and exceptions.
Consequently, rule‑based systems’ precision and adaptability may be limited.
In place of rule‑based methods, statistical machine translation (SMT) systems
have been developed. SMT uses statistical models that gather information from big
parallel corpora to learn translation patterns. SMT systems calculate the most likely
translations by looking at the frequency and co‑occurrence of words and phrases in
these corpora. SMT still has issues, while being a major improvement over rule‑based
approaches. SMT struggles to handle word sense disambiguation and uncommon or
unseen terms, is susceptible to the quality of the input data, and lacks contextual
comprehension.
The development of machine learning techniques, notably neural networks, revo‑
lutionized the area of machine translation by overcoming the drawbacks of rule‑based
and statistical approaches. Deep learning architectures are used by neural machine
translation (NMT) systems to recognize complex correlations and relationships in
language input.5 By directly learning from huge parallel corpora, certain architec‑
tures, such as recurrent neural networks (RNNs) and transformer models, make
end‑to‑end translation possible. By utilizing their capacity to record context, man‑
age long‑range dependencies, and generalize across several language pairs, NMT
systems have demonstrated significant improvement in delivering more accurate and
fluent translations.
Machine translation research and application have undergone considerable change
as a result of the introduction of machine learning techniques, notably neural machine
translation. Machine learning has paved the way for more reliable and context‑aware
translation systems by eschewing rigid rule‑based systems and statistical models
constrained by data scarcity.5 As a result, machine translation has advanced in terms
Data‑Centric AI Approaches for Machine Translation 57

FIGURE 4.1 Overview of machine translation systems.

of accuracy, idiomatic expression handling, and ability to adapt to various domains

and language combinations.
The limits of rule‑based and statistical approaches are the driving force for the use
of machine learning techniques in machine translation, which is highlighted in this
overview. In the parts that follow, we’ll go into greater detail on how machine learn‑
ing techniques, particularly neural networks, are used and how they affect machine
translation quality and performance.
58 Data-Centric Artificial Intelligence for Multidisciplinary Applications

4.3 MACHINE LEARNING TECHNIQUES

With new opportunities for enhancing translation accuracy and quality, machine
learning techniques have revolutionized machine translation. The core ideas and
algorithms used in machine learning for machine translation are explained in this
section, along with their advantages over earlier rule‑based methods. RNNs and
transformer models, two well‑known machine learning models used in machine
translation, are also examined.
Machine learning for machine translation is based on neural networks, computa‑
tional models inspired by the structure and function of the human brain. Artificial
neurons that are interconnected and arranged in layers make up neural networks. Deep
learning, a subset of machine learning, is the process of learning and extracting compli‑
cated patterns from data using deep neural networks, which include numerous layers.
Machine learning techniques, in contrast to conventional rule‑based methods, learn
patterns and correlations directly from data. The models may capture the fundamen‑
tal structure of language using this data‑driven methodology and generate predictions
based on observed patterns. Machine learning algorithms can find patterns in language
structures, word associations, and semantic linkages by using massive parallel corpora.
Machine learning techniques provide numerous benefits over conventional
rule‑based approaches in machine translation. In order to produce more accurate
and natural translations, machine learning models can capture the subtleties and
intricacies of language, such as idiomatic phrases and word sense disambiguation.
Additionally, machine learning approaches are more adaptable since they can pick
up new vocabulary from different language pairs and data from different domains.
A common class of machine learning models used in machine translation are
RNNs. By integrating feedback connections, RNNs are created to process sequential
data and keep track of prior inputs. This memory component enables RNNs to take
into account word dependencies and context, resulting in more accurate and contex‑
tually appropriate translations.
Transformer models, yet another important machine learning architecture, have
become more well‑known in recent years due to their outstanding translation capa‑
bilities. Transformer models use self‑attention techniques to identify long‑range
relationships in sentences, which enables the models to successfully align the infor‑
mation in the source and target languages during translation. This attention technique
enables improved word reordering management, increased fluency, and improved
translation quality.
In conclusion, machine learning methods have transformed machine translation
by enabling data‑driven learning and capturing the complexity of language, particu‑
larly neural networks like RNNs and transformer models. These approaches have a
number of advantages over traditional rule‑based approaches, including improved
accuracy, adaptability, and the ability to handle linguistic peculiarities.

4.4 PREPROCESSING AND FEATURE ENGINEERING

Preprocessing procedures and feature engineering methods are essential in the field
of machine translation for improving translation quality. In addition to concentrating
Data‑Centric AI Approaches for Machine Translation 59

on the necessary preprocessing stages, such as tokenization, sentence alignment, and

language model training, this section also explores feature engineering strategies that
help to improve translation results.
The segmentation of text into discrete tokens or words, or tokenization, is a cru‑
cial preparatory step in machine translation.2 Machine translation systems obtain a
detailed grasp of the input by breaking sentences down into tokens, which enables
better handling of linguistic structures and syntactic variations. Depending on the
language and particular needs of the translation process, different tokenization strat‑
egies may be used.
Another crucial preprocessing step called sentence alignment lines up related sen‑
tences in parallel corpora to achieve exact correspondences between source and target
languages.2 As a foundation for understanding translation patterns and aligning par‑
allel sentences during the translation process, this alignment is essential for training
machine translation models. Sentence alignment uses a variety of techniques, includ‑
ing the Smith‑Waterman algorithm and Hidden Markov Model‑based alignment.
Language models that accurately represent the statistical and syntactic patterns of a
particular language are trained using monolingual data. These linguistic models help
to maintain linguistic coherence and produce fluent translations. Depending on the
required level of sophistication and the available computer resources, language mod‑
els can be taught using approaches like n‑gram models, RNNs, or transformer models.
The quality of translations can be significantly increased by using feature engi‑
neering techniques. The semantic and contextual information of words is encoded
into low‑dimensional vector representations by word embeddings, also referred
to as dispersed word representations. In order to manage word meaning changes,
synonyms, and polysemy, these embeddings enable machine translation systems,
which eventually result in more accurate translations. The word embedding methods
Word2Vec, GloVe, and FastText are all well‑liked.
To improve translation quality, feature engineering also makes use of
language‑specific features. These features make use of linguistic details peculiar
to a certain language, such as part‑of‑speech tags, syntactic constructions, or mor‑
phological data.4 Such elements help machine translation models by adding more
linguistic information and contextual signals, enabling more accurate translations
and addressing language‑specific issues.
Machine translation systems can achieve notable gains in translation quality by
utilizing preprocessing stages like tokenization, sentence alignment, and language
model training, as well as feature engineering approaches like word embeddings and
language‑specific features.

4.5 ARCHITECTURES OF NEURAL NETWORKS

FOR MACHINE TRANSLATION
Refer to Figure 4.2 for the complete breakdown of the architectures of neural net‑
works in machine translation.
RNNs: Machine translation frequently employs a family of neural network
topologies called RNNs.3 By keeping a concealed state that preserves the context of
previously processed input, RNNs are made to process sequential data. RNNs may
60 Data-Centric Artificial Intelligence for Multidisciplinary Applications

FIGURE 4.2 Architectures of neural networks.

Data‑Centric AI Approaches for Machine Translation 61

produce a translation in the target language from a source language input sequence
in the context of machine translation.
Contextual dependencies are captured by RNNs by propagating data from one
time step to the next. The information about the previously viewed words and their
context is encoded in the RNN’s hidden state.
This enables the model to take into account the history of the input sequence
and produce translations that are logical and appropriate for the context. Standard
RNNs, however, experience the vanishing gradient problem, which might restrict
their use.
Longest Short‑Term Memory (LSTM) networks, a part of RNNs, were developed
to overcome this drawback.3 In order to more effectively capture long‑term depen‑
dence, LSTMs contain memory cells and gating mechanisms.
While the gating mechanisms regulate the flow of information, the memory
cells enable the model to selectively store an update information in the future time.
Compared to simple RNNs, LSTMs have a better capacity to grasp long‑term depen‑
dencies and increase translation accuracy.3
Transformer Models: Transformer models have become a ground‑breaking
machine translation architecture. Transformers rely on self‑attention techniques to
capture contextual dependencies as opposed to RNN‑based systems, which analyze
the input stream sequentially.
Using self‑attention, each word in the input sequence pays attention to every other
word to assess their relative value, is the fundamental concept of transformers. This
enables the model to provide translations while taking the sentence’s whole context
into account.
Transformers perform simultaneous processing of the full input sequence, paying
attention to various sections of the sequence to acquire data and make translation
judgments. Through effective context modeling and the capturing of dependencies
over the whole sequence, this parallel processing improves translation accuracy.
Transformers include a self‑attention mechanism that enables the model to gener‑
ate the translation while weighing the significance of various terms in the original
text.
Transformers can better match the source and destination language information,
resulting in more accurate translations, by paying attention to pertinent source lan‑
guage information.
Transformers also provide positional encoding to take into consideration the input
sequence’s sequential nature. The model can capture the word order and preserve
the sentence structure during translation, thanks to positional encoding which offers
information about the placement of each word in the sentence.
Role of Attention Mechanisms: The alignment of the information in the source
and destination languages during translation depends critically on attention pro‑
cesses. When producing translations, attention enables the model to concentrate on
particular segments of the input sequence, matching the pertinent source data with
the intended result.
The ability of the model to identify which words in the source sentence are most
pertinent to each word in the target phrase is provided by attention mechanisms in
machine translation.4
62 Data-Centric Artificial Intelligence for Multidisciplinary Applications
Attention mechanism in machine translation.
FIGURE 4.3
Data‑Centric AI Approaches for Machine Translation 63

By taking into account the context and meaning of the original words, this align‑
ment aids the model in producing appropriate translations.
In order to translate one word from the source sentence into another in the target
phrase, attention mechanisms compute attention weights for each word in the source
sentence.
Figure 4.3 gives us a brief idea about the role of the attention mechanism.
Machine translation models based on transformers have shown to be especially
good at using attention processes. Transformers include a self‑attention mechanism
that enables the model to understand the relationships between every word in the
input sequence, leading to more precise alignment and context‑aware translation.
The model is more versatile and adaptive to various translation tasks because of the
attention mechanism’s ability to manage variable‑length input and output sequences.
In summary, by retaining hidden states, utilizing memory cells and gating mecha‑
nisms, and utilizing attention processes, neural network designs like RNNs, LSTMs,
and transformers capture contextual dependencies in machine translation.5
These architectures have transformed the area of machine translation by making
it possible to create models that can provide translations that are more accurate and
contextually suitable.

4.6 TRAINING AND OPTIMIZATION

Training Process of Machine Translation Models: During the training phase of
machine translation models, machine learning techniques are employed to improve
the model parameters based on a training dataset.
The Following are the General Steps of the Training Process: The initial stage is
to prepare the training data, which consists of parallel corpora with source language
utterances and their corresponding target language translations. Usually, the data is
preprocessed by tokenizing the sentences, dealing with unfamiliar words, and per‑
forming the appropriate cleaning procedures.4
The next step is to initialize the model’s parameters. The weights and biases of
the network layers may be among the parameters, depending on the neural network
design being employed.
Forward Propagation: The model receives the input source sentence during train‑
ing and uses forward propagation to produce a translation in the target language. As
a result, an output is generated once the input has been processed through the layers
of the neural network.
Loss Computation: After comparing the produced and target translations, a loss
function is used to determine how different they are from one another. Machine trans‑
lation frequently employs the loss functions cross‑entropy loss and sequence‑level
loss.
Backward Propagation: To determine the model parameter gradients, the loss is
backpropagated across the network. In this stage, the overall loss contribution from
each parameter is calculated.
Model Parameter Update: Based on the computed gradients, optimization algo‑
rithms are utilized to update the model parameters. The goal is to reduce the loss
function as much as possible while enhancing translation quality.
64 Data-Centric Artificial Intelligence for Multidisciplinary Applications

FIGURE 4.4 Optimization of algorithm.

Iterative Optimization: For several training samples, Steps c to f are performed

iteratively to create epochs. A whole trip across the training dataset is represented
by an epoch. Depending on how quickly the model converges, there may be different
numbers of epochs.
Evaluation: To track the model’s development and avoid overfitting, the model’s
performance is assessed on a distinct validation set at various points during the train‑
ing phase. Using this evaluation, the best model can be chosen.
Optimization Algorithms: Stochastic Gradient Descent (SGD) is a popular machine
learning optimization algorithm. It modifies the model parameters by computing the
gradients on a randomly selected subset (mini‑batch) of the training data. Because of
this random selection, the method is computationally efficient and less prone to data
noise. SGD updates are carried out in little stages, shifting the parameters to mini‑
mize the loss function.
Refer to Figure 4.4 for the understanding of the algorithm optimization process.
Adam (Adaptive Moment Estimation): Adam is an extension of SGD that modi‑
fies the learning rate for each model parameter in accordance with predictions of
the first and second moments of the gradients. It provides faster convergence and
better handling of sparse gradients by combining the advantages of the AdaGrad and
RMSprop techniques.
The essential phases of the Adam optimization method are as follows:

i. Initialization: For each model parameter, Adam initializes the first and
second moment variables (m and v) to zero.
ii. Gradient Computation: Based on the small batch of training examples,
the gradients of the model parameters are computed during the forward
propagation and backpropagation steps.
iii. Moment Updates: Adam determines the exponential moving averages of
the first (m) and second (v) moments of the gradients. The decay rates used
to calculate these moving averages are 1 and 2, respectively.
iv. Bias Correction: The moving average estimates may be biased toward
zero because they are initialized with zeros, especially in the first training
Data‑Centric AI Approaches for Machine Translation 65

iterations. By modifying the estimates with a bias correction term, Adam

corrects for bias.
v. Adam changes the model parameters using the bias‑corrected moment esti‑
mates and a learning rate after computing the estimates. The formula for
( )
updating a parameter is as follows: θ = θ − (α * m ) ÷ v + ε , which is a
minor constant introduced for numerical stability.

Based on the first and second moment estimations, Adam modifies the learning rate
for each parameter, causing the model to converge more quickly and handle various
gradients effectively.

4.7 EVALUATION AND METRICS

Evaluation Process for Machine Translation Systems: To gauge the effectiveness of
machine translation systems and direct improvement efforts, quality evaluation is
essential. Both manual and automatic measures are used in the machine translation
evaluation process:

a. Human Assessment: In human assessment, translations produced by

machine translation systems are evaluated by human judges. These judges
frequently speak the target language fluently or bilingually. They evaluate
the machine‑generated translations in comparison to the reference transla‑
tions and assign scores or rankings based on a number of factors, including
fluency, sufficiency, grammaticality, and overall quality. The linguistic pro‑
ficiency and naturalness of the translations are crucial insights gained via
human inspection.
b. Automated Metrics: Automated metrics are impersonal measurements
created to evaluate the accuracy of machine translations without the need
for human interaction. These metrics measure the machine‑translated texts
against one or more reference translations and compare the results to pre‑
determined standards. Large amounts of translated material can be evalu‑
ated quickly and effectively with the use of automated metrics. They are
commonly used in machine translation system research, development, and
benchmarking.

4.7.1 Commonly Used Evaluation Metrics

BLEU (Bilingual Evaluation Understudy): BLEU is a well‑liked assessment met‑
ric that is frequently employed in the study and improvement of machine translation.
It is built on the idea of n‑gram accuracy, which quantifies how closely the reference
translations and machine translations match in terms of word sequences.
By measuring the number of matching n‑grams in the machine‑generated trans‑
lation and the reference translations, BLEU determines the precision of various
n‑gram orders (usually unigrams, bigrams, trigrams, and occasionally higher orders).
Additionally, it penalizes for producing more n‑grams than the references contain.
A weighted geometric mean is then used to aggregate the precision scores.
66 Data-Centric Artificial Intelligence for Multidisciplinary Applications

BLEU has some restrictions even if it is now a de facto standard. Fluency, ade‑
quacy, and word order are not taken into account by BLEU, which primarily con‑
centrates on lexical similarity. Because it is sensitive to superficial similarities,
translations that might not accurately convey the intended meaning often receive
high marks. Furthermore, BLEU is domain‑dependent, which means that its perfor‑
mance may vary across various text domains.
Metric for Evaluation of Translation with Explicit Ordering or METEOR:
Another popular evaluation metric, METEOR, seeks to capture many facets of trans‑
lation quality. It takes into account a variety of matching factors, such as unigrams,
stemming, synonymy, word order, and more.
Considering the word alignment between the reference translations and the
machine‑generated translation, METEOR determines precision and recall values. In
addition to a number of matching and penalty systems, the ultimate score is calcu‑
lated using a harmonic mean of recall and precision.
METEOR has the advantage of handling word‑order variations, accounting for
synonyms, and stemming.4 It can distinguish between variations in words’ surface
forms and offer more precise alignments. It has been demonstrated that METEOR
performs well across a variety of language pairs and text domains. It does, however,
have some restrictions. METEOR largely depends on the caliber of available linguis‑
tic resources, including word alignments, stemmers, and synonyms. Because of its
complicated scoring system, it is challenging to evaluate and comprehend the precise
contribution of each component.
TER (Translation Edit Rate): The Edit Distance between the Machine‑Generated
Translation and the Reference Translation (TER) is a measurement used in
­evaluation. It determines how many edits – including additions, subtractions, and
substitutions – are necessary to turn the machine‑generated translation into the refer‑
ence translation.
TER concentrates on the surface‑level alterations and offers a direct measurement
of translation similarity.4 Since it is linguistically neutral, it can be applied to other
language combinations. A text that has been heavily edited or post‑edited, where the
emphasis is on the number of revisions done, can be evaluated well using TER.
TER does, however, have some restrictions. It does not take semantic equiva‑
lence or fluency into account and rather concentrates on the number of modifications.
It may penalize translators for legitimate paraphrases or rearranging sentences to
increase readability or flow. Additionally, because TER scores cannot fully capture
the whole meaning or appropriateness of the translation, they may not be able to
provide a comprehensive assessment of translation quality.

4.7.2 Strengths and Weaknesses of Evaluation Metrics

Each evaluation metric has its own advantages and disadvantages, and the degree
to which they are correlated with human judgments depends on a number of vari‑
ables. Here are some important details about these evaluation metrics’ advantages
and disadvantages:
Data‑Centric AI Approaches for Machine Translation 67

4.7.2.1 BLEU
a. Strengths: BLEU is widely used, simple to implement, and effective com‑
putationally. It offers an immediate evaluation of translation quality and is
simple to understand. For specific language pairs and text domains, BLEU
scores have a fair amount of agreement with human assessments.
b. Weakness: BLEU’s emphasis on lexical similarity and n‑gram accuracy
can produce high ratings for translations that may not accurately convey
the intended meaning. Fluency, sufficiency, and word order – which are
essential for assessing translation quality – are not taken into account.
BLEU is susceptible to manipulation and is also perceptive to superficial
similarities.

4.7.2.2 METEOR
Strengths: METEOR provides a more thorough assessment of translation
quality by taking many linguistic factors like word order, stemming, and
synonymy into account. It performs effectively in a variety of language and
domain combinations. In comparison to BLEU, METEOR can tolerate vari‑
ances in surface shapes and offer superior alignments.
Limitations: The effectiveness of METEOR significantly depends on
the availability and caliber of linguistic resources like stemmers, synonyms,
and word alignments. Its scoring system is intricate and may be difficult to
understand. When linguistic resources are scarce or poorly matched to the
target language, METEOR scores may not always match human judgments
exactly.

4.7.2.3 TER
a. Strengths: TER provides an accurate comparison of similarity by mea‑
suring the edit distance between translations. It is not linguistically reliant
and is less prone to fluency and adequacy problems. Since TER can detect
significant translational changes, it is helpful for assessing heavily edited or
post‑edited text.
b. Weaknesses: TER does not take semantic equivalence into account and
rather concentrates on the number of revisions. It may penalize translators
for legitimate paraphrases or rearranging sentences to increase readability
or flow. As TER scores do not fully capture the whole meaning or appro‑
priateness of the translation, they may not be able to provide a thorough
assessment of translation quality.

It’s critical to realize that while these evaluation criteria provide quantitative assess‑
ments of translation quality, they do not fully account for human perception. In order
to gain a thorough understanding of machine translation system performance, they
should be utilized in conjunction with human review and other qualitative analyses.
Additionally, depending on the language pair, text domain, and particular translation
issues, the applicability and efficacy of these indicators may change.
68 Data-Centric Artificial Intelligence for Multidisciplinary Applications

4.8 METHODOLOGY
Machine learning techniques play a crucial role in machine translation since they
have completely changed how we approach language translation tasks. The process
for using machine learning in machine translation is described below, and refer to
Figure 4.5 for brief information about the methodology:

1. Issue Description: Define the precise translation task, such as speech‑to‑text,

image‑to‑text, or text translation from one language to another.
2. Data Gathering and Preparation: A mass of a sizable, parallel corpus of
text or information in the target and source languages. By removing noise,
special characters, and unnecessary information, clean up and preprocess
the data. To construct a vocabulary, tokenize the text into words or subword
units (like BPE or SentencePiece).
3. Extracting Features: You can encode words or subword units as numeri‑
cal vectors by using methods like word embeddings (Word2Vec, GloVe,
FastText), contextual embeddings (BERT, GPT, ELMo), or subword embed‑
dings (WordPiece, Byte‑Pair Encoding).
4. Model Choice: Select a machine translation‑friendly machine learning
model. Typical options include phrase‑based SMT models are one type of
SMT model. Models for NMT, such as attention‑based sequence‑to‑sequence
models. Models built on transformers (such BERT, GPT, and T5) have been
honed for translation jobs.
5. Post‑Processing: Use post‑processing methods to improve the output of the
model, such as capitalization, punctuation, and fluency.
6. Deployment: Use the trained machine translation model in applications and
systems as an API, a service, or an integration where translation is needed.
7. Constant Development: Continually evaluate model performance in
real‑world circumstances and get user input. To enhance translation quality,
repeat the training process with newer data and retrained models. To man‑
age several languages, think about developing or improving multilingual
models. Create specialized models for translations in fields like law, medi‑
cine, and technology. Continuously gather user feedback and monitor model
performance in real‑world scenarios.

4.9 ADVANCED TECHNIQUES AND RECENT DEVELOPMENTS

Advanced Techniques in Machine Translation: With the advent of several
advanced methods, machine translation has made great strides in recent times. Here
are some notable methods that have helped machine translation progress:

1. NMT models, especially those based on deep learning architectures such as

RNNs and transformers, have revolutionized machine translation.2
2. These models effectively capture context and long‑term interdependencies
while learning to translate by analyzing large amounts of multilingual tex‑
tual data.2 Figure 4.6 gives us an example of ​​the advanced techniques in the
machine translation process.
Data‑Centric AI Approaches for Machine Translation 69

FIGURE 4.5 Methodology diagram for machine translation.

70 Data-Centric Artificial Intelligence for Multidisciplinary Applications

FIGURE 4.6 Machine translation and confidentiality.

Transition Learning: Transfer learning techniques make it possible to pre‑train

language models on large datasets and refine them for specific translation tasks. By
using knowledge gained from huge volumes of data, this method helps to improve
translation quality.
Attention Mechanism: Attention mechanisms improve the performance of
machine translation models by allowing them to focus on relevant phrases from the
source text during translation. Attention processes facilitate the copying of important
information and improve the overall quality of the translation by assigning different
weights to certain words or phrases.
Many machine translation methods use subword units, including subword embed‑
ding or byte pair encoding (BPE), rather than translating words at the word level.2
Secondary word segmentation increases translation accuracy, especially for mor‑
phologically rich languages, by dealing with uncommon words, out‑of‑vocabulary
terms, and rare words. Zero‑Shot Translation and Multi‑Language Translation:
Multilingual translation models can handle multiple languages at the same time,
allowing efficient translation between different language pairs. This skill is extended
by Zero Hit Translation, which allows translation between language pairs that are not
directly visible during training.
Machine translation models can be improved using reinforcement learning strat‑
egies, such as reinforcement learning from human feedback. These methods aim
to improve translation accuracy and fluency by using human‑generated feedback to
reinforce desired translation results.
Unattended Machine Translation: Without using a parallel corpus, unsuper‑
vised machine translation seeks to train translation models. Low‑resource‑intensive
languages can
​​ be translated because unsupervised algorithms can train themselves
to translate without concatenated versions using monolingual data from the source
and target languages.
Data‑Centric AI Approaches for Machine Translation 71

FIGURE 4.7 Neural machine translation with sequence‑to‑sequence RNN.

1. Recent Developments in Neural Machine Translation: NMT has

recently made advances in unsupervised techniques and pre‑trained lan‑
guage models, such as BERT (Bidirectional Encoder Representation from
Transformers).5 Here are some notable developments, and Figure 4.7 gives
us an idea of ​​a Sequential to Sequential Neural Machine Translation RNN.
2. Pre‑Trained Language Models for NMT: A pre‑trained language model
that has been successfully applied to improve the performance of NMT is
BERT. The pre‑trained BERT integration, which contains detailed back‑
ground information, can be used to initialize the NMT model instead of
starting from scratch. This transfer learning strategy helps NMT models bet‑
ter understand the source language and produce more accurate translations.
3. Contextual Embedding: By using contextual embeddings provided
by models such as BERT, the representation of words in NMT has been
improved. In situations where it is important to define a word, these embed‑
dings capture the context and meaning of the word based on the terms sur‑
rounding the word, resulting in more accurate translations.
4. Multilingual Language Model: To support NMT in multiple languages,
multilingual language models, such as XLM (Multilingual Language
Model), have been developed.6 These models are pre‑trained on huge
amounts of multilingual data, facilitating knowledge transfer between lan‑
guages ​​and improving translation quality for under‑resourced language
pairs.
5. Unattended Machine Translation (NMT): NMT models are trained using
unsupervised machine translation techniques that do not rely on parallel
corpus. These methods use monolingual data from the source and target
languages and use tools such as denoising encoders and automatic decom‑
pilers to learn to translate without concatenation problems. Unsupervised
NMT has shown promising results, especially for low‑resource languages ​​
where parallel training material is inadequate.
72 Data-Centric Artificial Intelligence for Multidisciplinary Applications

6. Multilingual NMT: Using a single model, multilingual NMT models can

translate between multiple languages. These models take advantage of
shared representations between languages, improving translation quality
and maximizing resource usage.

4.10 CHALLENGES AND FUTURE DIRECTIONS

1. Challenges: The accuracy and quality of machine translation are hampered
by a number of issues. Here are some pressing problems and their solutions,
which are also represented by the flowchart in Figure 4.8:
• Languages with Few Resources: The scarcity of parallel training data
in low‑resource languages makes it difficult to develop reliable transla‑
tion models. The researchers looked at methods such as unsupervised
machine translation, to solve this problem, training translation models
using monolingual data from source and target languages. By applying
lessons from high‑resource languages ​​and using multilingual models,
translation quality for low‑resource languages ​​can also be improved.

FIGURE 4.8 Challenges and solution.

Data‑Centric AI Approaches for Machine Translation 73

• Idiom Expressions: Idioms are words or phrases that can be taken

literally but also figuratively.4 Since their meanings may not translate
word for word correctly, it can be difficult to translate these expres‑
sions effectively. Knowledge of context is required to manage idiomatic
languages, and attention‑processing neural machine translation models
have shown progress in capturing these nuances. Idioms can also be
translated using strategies such as phrase‑based translation and the use
of bilingual dictionaries or phrase boards.
• Domain Adaptation: In specific fields, such as legal or medical transla‑
tion, where specialized terminology and context are common, machine
translation models often have difficulty performing well. Domain adap‑
tation techniques can be used to improve or train models specific to
certain domains.
• Rare and Extra‑Lexical Words: When machine translation encounters
rare or out‑of‑vocabulary (OOV) words that are not seen during train‑
ing, problems can arise.4 By dividing OOV words into sub‑word units,
sub‑word segmentation algorithms such as BPE or character‑based
models can process them. These methods minimize the influence of
OOV terms on translation quality and allow the translation model to
handle rare words more efficiently.
• Polysemy and Ambiguity: When a term or phrase has multiple inter‑
pretations, the result is ambiguity. The phenomenon of a word having
many similar meanings is called polysemy. For a translation to be hon‑
est, it is essential to remove ambiguity and polysemy. Attention‑based
neural machine translation models can help capture contextual informa‑
tion to distinguish words or phrases based on context.
2. Future Directions: Current machine translation research focuses on
a number of intriguing directions and emerging trends. Here are some
examples:
• Context Data: An important area of ​​research in machine translation
is the incorporation of contextual data. More accurate and more con‑
textual translations can be generated by models that take into account
the surrounding context of the source sentence. Context‑aware methods,
such as those that use context information at the conversation or docu‑
ment level, are being explored in an effort to improve translation quality
and consistency.
• Translations of Rare Words: For accurate and fluent translation, it is
essential to improve the translation of rare terms. Using sub‑word units,
morphological analysis, or context‑based embedding, researchers are
working on methods to better manage rare and infrequent words. The
meaning and usage of uncommon words can be accurately captured
by machine translation models, allowing them to generate more and
maintain fluency.
• Improving the consistency and fluency of machine translation results
is an important area of research.
​​ More natural‑sounding, stylistically
consistent, and consistent translations are produced using methods
74 Data-Centric Artificial Intelligence for Multidisciplinary Applications

such as neural text‑generation modeling, reinforcement learning, and

iterative refinement. Ensuring that translated documents read smoothly
and naturally is key to improving user experience and driving machine
translation adoption.
• Multi‑Modal Translation: Research focuses on integrating visual or
audio information into machine translation models in response to the
evolution of multimodal data (such as text accompanied by images or
sounds). Multimodal translation attempts to handle translation tasks
that require multiple methods and produce translations that are more
contextual by including visual or spoken cues.
• Improvements in Neural Architecture and Training: To advance
machine translation, researchers are working on improvements in
neural architecture and training methods. The goal of methods such
as knowledge distillation, multitasking learning, and model compres‑
sion is to improve performance without sacrificing translation quality,
model size, or efficiency. In addition, new designs such as transformer
variants and sparse neural networks are being explored to overcome the
limitations of translation efficiency and accuracy. Ongoing initiatives
to improve translation quality, manage specific constraints, and cre‑
ate more accessible and reliable translation systems are exemplified by
these evolving trends in machine translation and current research direc‑
tions. Machine translation has the potential to be an even more power‑
ful tool for breaking down language barriers and promoting effective
cross‑language communication.

4.11 CONCLUSION
The improvement in the quality, fluency, and accuracy of machine translations is
mainly due to machine learning techniques. In terms of pattern recognition, trans‑
lation models can be trained using machine learning methods to provide transla‑
tions that are very similar to human‑generated translations, understand context, and
make decisions. Here are some important functions of a machine learning strategy
in machine translation:
Pattern Recognition: Machine learning techniques allow translation models to
learn from large amounts of multilingual textual data and identify patterns that con‑
tribute to accurate translations. The lexical, syntactic, and semantic patterns in the
source language can be recognized by the models, which can then be used to gener‑
ate equivalent translations in the target language.
Contextual Understanding: Translation models can understand the context of
the source sentence more efficiently during translation through machine learning
techniques such as neural networks. To manage idioms, distinguish words or phrases,
and provide linguistically consistent and contextual translations, models can use con‑
textual information.
Learn from Data: To discover language matches between source and target lan‑
guages, machine translation models are mainly data‑driven. Models can be trained on
huge parallel corpus using supervised learning methods, where the source sentence
Data‑Centric AI Approaches for Machine Translation 75

and its corresponding translation are synchronized. Models can learn translation pat‑
terns using this data‑driven method, generalize from versions, and improve transla‑
tion quality over time.
Generalizability and Adaptability: Machine learning techniques make it easy
to adapt translation models to other languages, domains, and styles. To support addi‑
tional language pairs or specialized domains, models that have been trained on a
language pair or domain can be adapted or extended. Machine translation systems
can accommodate many translation requirements due to their flexibility, which also
helps them perform better over time.
Machine learning algorithms are excellent at collecting complex patterns and
dealing with language differences. Due to the ambiguities, idiomatic expressions,
and variations that exist between languages, machine translation is inherently dif‑
ficult. Rule‑based systems often have difficulty successfully handling such complexi‑
ties, while machine learning models can learn from huge volumes of data and adapt
to that complexity.
Data‑driven approach: Machine learning techniques rely on data‑driven
approaches and make translation decisions by learning from large volumes in paral‑
lel. Models can generalize from examples using this data‑centric approach, adapting
to different language and domain pairs, and improving over time. On the other hand,
rule‑based systems are less scalable and flexible because they require humans to
build and maintain rules significantly.
Generalizability and Adaptability: By adapting or extending their training
on specific data, machine learning models can easily generalize to new language,
domain, or style pairs. Machine translation systems can now adapt to a wide vari‑
ety of translation requirements without having to manually build or modify rules.
Traditional rule‑based methods have limited extensibility and generalization, as
every single translation job requires extensive rule engineering and maintenance.
Transformer model and RNN are two machine learning models commonly used
in machine translation. RNNs such as LSTM or closed‑repetition units are com‑
monly used in machine translation systems using a sequence‑by‑sequence model.
On the other hand, transforming toys have become a powerful alternative.
The foundational article “Attention is all you need” introduced the design of the
Transformer, which uses self‑attention techniques to capture dependencies through
the input string. Using this attention mechanism, Transformer models can pay atten‑
tion to relevant source words during translation, increase translation accuracy, and
handle long‑range dependencies more efficiently.

REFERENCES
[1] Brown, Peter F., John Cocke, Stephen A. Della Pietra, Vincent J. Della Pietra, Frederick
Jelinek, John Lafferty, Robert L. Mercer, and Paul S. Roossin. “A statistical approach to
machine translation.” Computational Linguistics 16, no. 2 (1990): 79–85.
[2] Zhao, Bei, and Wei Gao. “Machine Learning Based Text Classification Technology.”
In: 2022 IEEE 2nd International Conference on Mobile Networks and Wireless
Communications (ICMNWC), IEEE, 2022, pp. 1–5.
[3] Sutskever, Ilya, Oriol Vinyals, and Quoc V. Le. “Sequence to sequence learning with
neural networks.” Advances in Neural Information Processing Systems 27 (2014).
76 Data-Centric Artificial Intelligence for Multidisciplinary Applications

[4] Sennrich, Rico, Barry Haddow, and Alexandra Birch. “Neural machine translation of
rare words with subword units.” (2015). arXiv preprint arXiv:1508.07909.
[5] Benkov, Lucia. “Neural machine translation as a novel approach to machine trans‑
lation.” In: DIVAI 2020 The 13th International Scientific Conference on Distance
Learning in Applied Informatics, 2020, pp. 499–509.
[6] Bkassiny, Mario, Yang Li, and Sudharman K. Jayaweera. “A survey on machine‑learning
techniques in cognitive radios.” IEEE Communications Surveys & Tutorials 15, no. 3
(2012): 1136–1159.
Section II
Data‑Centric AI in Healthcare
and Agriculture
 5 Case Study Medical
Images Analysis and
Classification with
Data-Centric Approach
Namrata N. Wasatkar and Pranali G. Chavhan

5.1 INTRODUCTION: BACKGROUND AND DRIVING FORCES

Medical image analysis plays a crucial role in modern healthcare, enabling accu‑
rate diagnosis, treatment planning, and monitoring of various medical conditions.
With the advent of advanced technologies and the availability of large‑scale medical
image datasets, there is a growing need to develop efficient and reliable methods for
analyzing and interpreting these images [1]. The data‑centric approach in medical
image analysis focuses on acquiring high‑quality datasets, pre‑processing the images
to ensure accuracy, extracting meaningful features, and utilizing robust classifica‑
tion models [2,3]. By prioritizing data quality, this approach aims to optimize the
performance and reliability of the analysis and classification process. Through the
integration of machine learning, deep learning, and image processing techniques,
the data‑centric approach holds tremendous promise in enhancing medical imaging
practices, improving healthcare decision‑making, and ultimately benefiting patients
worldwide [4]. Medical Images Analysis and Classification with a data‑centric
approach is a comprehensive and innovative field that leverages advanced techniques
to extract valuable insights from medical images [4].
Medical image classification plays a critical role in modern healthcare, enabling
accurate diagnosis, treatment planning, and patient management. With the rapid
advancement of imaging technologies and the exponential growth of medical
image data, there is an increasing need to develop efficient and reliable methods
for classifying these images effectively [5]. The data‑centric approach in medical
image classification emphasizes the significance of acquiring high‑quality datasets,
pre‑processing the images to enhance their quality and remove noise or artifacts,
extracting relevant and discriminative features, and utilizing robust classification
models [6].
By adopting a data‑centric approach, healthcare practitioners and researchers
can ensure that the classification process is optimized and based on accurate and
representative data. The quality and diversity of the dataset are paramount to the
success of the classification task, as it enables the model to learn and generalize pat‑
terns effectively. Acquiring a diverse dataset that covers various medical conditions,

DOI: 10.1201/9781003461500-7 79
80 Data-Centric Artificial Intelligence for Multidisciplinary Applications

imaging modalities, and patient demographics is essential to train a robust and

reliable classifier [7]. Once the dataset is obtained, pre‑processing techniques are
applied to enhance the image quality and remove any inconsistencies or imperfec‑
tions that may affect the classification accuracy. Pre‑processing steps may involve
resizing images, normalizing pixel values, applying filters, and handling missing
data [8]. These steps ensure that the input images are in a suitable format for subse‑
quent analysis and classification.
Feature extraction is a crucial step in medical image classification, as it involves
converting the raw image data into meaningful and informative representations.
Different techniques can be employed, ranging from traditional statistical features
to more advanced deep learning‑based approaches. These features aim to capture
distinctive characteristics and patterns that are relevant to the classification task.
By extracting discriminative features, the model can differentiate between different
medical conditions accurately [9].
The selection of an appropriate classification model is another key aspect of the
data‑centric approach. Various models can be utilized, including traditional machine
learning algorithms such as support vector machines, random forests, or more
advanced deep learning architectures like convolutional neural networks (CNNs) [10].
The choice of the model depends on the complexity of the classification task, the avail‑
able computational resources, and the interpretability of the results. It is crucial to
select a model that strikes a balance between accuracy and computational efficiency.
Once the model is selected, it undergoes a training process using the pre‑­processed
dataset. During training, the model learns to classify medical images based on the
extracted features. The model’s hyperparameters, such as learning rate, regulariza‑
tion, and architecture, are fine‑tuned to optimize its performance. The trained model
is then evaluated using appropriate metrics such as accuracy, precision, recall, and F1
score to assess its classification performance [11].
By leveraging the data‑centric approach, medical image classification has the
potential to significantly improve patient care. Accurate and reliable classification
of medical images enables timely and appropriate treatment decisions, reduces the
chances of misdiagnosis, and enhances overall healthcare outcomes. Furthermore,
the continuous refinement and improvement of the classification process based on
feedback from healthcare professionals and researchers can lead to advancements in
medical imaging practices and contribute to the development of innovative diagnos‑
tic tools.
In supposition, the data‑centric approach in medical image classification priori‑
tizes the acquisition of high‑quality datasets, pre‑processing of images, extraction of
meaningful features, and utilization of robust classification models [12]. By lever‑
aging advanced techniques and placing an emphasis on data quality, this approach
has the potential to revolutionize medical image classification, leading to improved
patient care and advancing the field of medical imaging.
By employing a data‑centric approach, this methodology focuses on acquiring
high‑quality datasets, pre‑processing them to enhance their accuracy, extracting rel‑
evant features, and applying suitable classification models. This approach enables
healthcare professionals and researchers to efficiently analyze and classify medical
images, leading to improved diagnostic accuracy, treatment planning, and patient
Case Study Medical Images 81

outcomes [13]. By combining cutting‑edge technologies like machine learning, deep

learning, and image processing, this approach holds tremendous potential in revolu‑
tionizing medical imaging practices and empowering healthcare practitioners with
powerful tools for diagnosis and decision‑making.

5.2 LITERATURE SURVEY

1. Bullock, J., Cuesta‑Lázaro, C. and Quera‑Bofarull, A., 2019, March. XNet:
a CNN implementation for medical X‑ray image segmentation suitable for
small datasets.
This paper presents a case study on the application of a data‑centric approach
to analyze and classify X‑ray images for diagnosing lung diseases. The authors
explore feature extraction techniques, feature selection methods, and model
selection for accurate classification. They demonstrate the effectiveness of the
data‑centric approach in achieving high‑accuracy rates for disease classifica‑
tion. With this approach, it provides overall accuracy of 92%, an F1 score of
0.92, and an AUC of 0.98, surpassing classical image processing techniques,
such as clustering and entropy‑based methods, while improving upon the out‑
put of existing neural networks used for segmentation in non‑medical contexts.
2. Wang, J., Zhu, H., Wang, S.H. and Zhang, Y.D., 2021. A review of deep
learning on medical image analysis. Mobile Networks and Applications, 26,
pp.351–380.
This comprehensive review discusses the use of deep learning techniques in
medical image analysis. The authors present case studies on the application of
deep learning models for tasks such as tumor detection, lesion segmentation,
and disease classification in medical imaging.
3. Terzi, R., Azginoglu, N. and Terzi, D.S., 2022. False positive repression:
Data centric pipeline for object detection in brain MRI. Concurrency and
Computation: Practice and Experience, 34(20), p.e6821.
This study presents a data‑centric approach for brain tumor classifica‑
tion using MRI images. The authors present a case study on brain tumor
­classification, comparing the performance of different feature selection and
classification algorithms, demonstrating the effectiveness of the data‑centric
approach in accurate tumor classification. The authors suggested pipeline out‑
performed the traditional pipeline by up to 18% on the Gazi Brains 2020 data‑
set and up to 24% on the BraTS 2020 dataset for mean specificity value without
affecting the sensitivity measure significantly. This suggests that the suggested
pipeline minimizes false positive rates caused by bias in real‑world applica‑
tions and can assist in reducing expert workload.
4. Swedhaasri, M., Parekh, T. and Sharma, A., 2021, August. A Multi‑Stage
Deep Transfer Learning Method for Classification of Diabetic Retinopathy
in Retinal Images. In 2021 Second International Conference on Electronics
and Sustainable Communication Systems (ICESC) (pp. 1143–1149). IEEE.
This paper explores the application of a data‑centric approach for retinal
image analysis and the detection of diabetic retinopathy. The authors discuss
82 Data-Centric Artificial Intelligence for Multidisciplinary Applications

various image Pre‑processing techniques, feature extraction methods, and

classification models. They present a case study on diabetic retinopathy detec‑
tion using retinal fundus images, demonstrating the effectiveness of the data‑­
centric approach in accurate disease classification. The authors proposed study
effort proposes a multi‑stage deep transfer learning approach that includes
labeling using similar datasets, which is fundamentally distinct. The approach
proposed in this study may be used as a screening tool for early identification
of DR, with a sensitivity and specificity of 0.99 and a quadratic weighted kappa
grade of 0.925466 of the blindness detection dataset.
This paper provides insights into the application of a data‑centric approach
in the analysis and classification of medical images for different healthcare
scenarios. They cover various imaging modalities, pre‑processing techniques,
feature engineering methods, and classification models, highlighting the ben‑
efits and effectiveness of the data‑centric approach in improving accuracy and
performance in medical image analysis.
5. Chen, C.C., DaPonte, J.S. and Fox, M.D., 1989. Fractal feature analysis and
classification in medical imaging. IEEE transactions on medical imaging,
8(2), pp.133–142.
In this article, there are two applications: (1) classification and (2) edge
enhancement and detection. This estimate approach is used to define a nor‑
malized fractional Brownian motion feature vector for classification purposes.
It represented the normalized average absolute intensity difference between
pixel pairs on a scaled surface. The feature vector represents the statistical
properties of the medial image surface with a small number of data items and
is insensitive to linear intensity adjustment. A modified picture is created for
edge enhancement and detection applications by computing the fractal dimen‑
sion of each pixel throughout the whole medical image.
6. Mohan, G. and Subashini, M.M., 2018. MRI based medical image analysis:
Survey on brain tumor grade classification. Biomedical Signal Processing
and Control, 39, pp.139–161.
This is where digital image processing techniques and machine learning
come into play, assisting with further diagnosis, therapy, before and post‑­
surgical procedures, and bridging the gap between the radiologist and the
computer. These hybrid approaches give radiologists a second viewpoint and
support in comprehending medical pictures, hence boosting diagnosis accu‑
racy. This article will look back at current advances in segmentation and clas‑
sification in tumor‑infected human brain MR images, with a focus on gliomas,
including astrocytoma. The approaches for tumor extraction and grading that
may be integrated into routine clinical imaging protocols are described.

5.3 IMPLEMENTATION‑MEDICAL IMAGE ANALYSES

Step 1: Dataset Exploration and Pre‑processing
Exploring and pre‑processing a dataset with a data‑centric approach
involves understanding the data, handling missing values, dealing with out‑
liers, and transforming the data to make it suitable for analysis [14].
Case Study Medical Images 83

Step 2: Data Visualization

Data visualization plays a crucial role in understanding and communi‑
cating insights from data. The data visualization process should be driven
by the specific characteristics and requirements of data, as well as the objec‑
tives of the analysis. By adopting a data‑centric approach, data visualizations
effectively communicate insights and support data‑driven decision‑making.
Step 3: Feature Engineering
Feature engineering is an iterative process, and it requires a combina‑
tion of analytical skills, domain knowledge, and creativity. By adopting a
data‑centric approach, we can uncover hidden patterns, improve model per‑
formance, and gain deeper insights from your data [15].
Step 4: Feature Selection
Feature selection is the process of identifying and selecting the most
relevant and informative features from a dataset to improve model perfor‑
mance, reduce overfitting, and enhance interpretability. When approach‑
ing feature selection with a data‑centric perspective, consider the following
points:
• Define the Evaluation Metric
• Perform Initial Feature Exploration
• Filter Methods
• Wrapper Methods: Utilize wrapper methods that evaluate feature sub‑
sets by training models and measuring their performance [16, 17, 20].
• Embedded Methods: Leverage embedded methods that perform fea‑
ture selection as part of the model training process [21].
• Model‑Based Evaluation: Train your model using selected features
and evaluate its performance using the defined evaluation metric [22].
• Iterative Refinement: Iterate the feature selection process by fine‑­
tuning the selection criteria, exploring different methods, or incorporat‑
ing domain knowledge.
• Consider Dimensionality Reduction: If dealing with high‑dimen‑
sional datasets, consider dimensionality reduction techniques (e.g.,
Principal Component Analysis or t‑SNE) to transform the data into a
lower‑dimensional space while retaining important information.
Feature selection is a trade‑off between model simplicity, interpret‑
ability, and predictive performance. A data‑centric approach involves
thoroughly evaluating different feature selection techniques, consid‑
ering the specific requirements of your modeling task, and iteratively
refining your feature selection process based on feedback and perfor‑
mance evaluation.
Step 5: Model Selection and Hyperparameter Tuning
It involves choosing the appropriate algorithm and optimizing its hyper‑
parameters to achieve the best performance. When approaching model
selection and hyperparameter tuning with a data‑centric perspective. The
data‑centric approach involves an iterative process that considers the char‑
acteristics of the data, the specific modeling task, and the evaluation metrics
[18]. By systematically exploring different models, tuning hyperparameters,
84 Data-Centric Artificial Intelligence for Multidisciplinary Applications

and evaluating performance, we can select the best model and optimize its
performance for our data.
Step 6: Model Evaluation
Model evaluation is a crucial step in the data modeling process to assess
the performance and effectiveness of a trained model. Model evaluation
is an ongoing process that requires careful consideration of the evaluation
metrics, performance strategies, and the specific goals of your analysis. By
adopting a data‑centric approach, you can gain a deeper understanding of
your model’s performance, make informed decisions, and refine your mod‑
eling techniques to optimize results.

5.4 BREAST CANCER WISCONSIN DATASET

Implementing data‑centric approach. However, we provided a code template that out‑
lines the steps of the data‑centric approach for the Breast Cancer Wisconsin dataset
[19]. This code may need to modify and expand specific sections based on specific
requirements and insights during the analysis.
The code as per Figures 5.1 and 5.2 provides a simplified example, and there
are many more aspects to consider, such as hyperparameter tuning, cross‑valida‑
tion, and more in‑depth feature engineering and analysis. This code snippet serves
as a starting point to demonstrate the data‑centric approach for the Breast Cancer
Wisconsin dataset.

FIGURE 5.1 Code for implementation of regression without preprocessing data.

Case Study Medical Images 85

FIGURE 5.2 Code for implementation of regression with preprocessing of data.

5.5 COMPARATIVE OF MODEL CENTRIC APPROACH

In the model‑centric approach, the primary focus is on building and optimizing
­models, often with less emphasis on thorough data pre‑processing and analysis. In
the model‑centric approach, the above code directly applies the model to the dataset
without extensive data preprocessing. While this approach is quicker, it might lead
to suboptimal results if the data is not properly prepared or if feature engineering is
necessary.
Both approaches have their place depending on the available resources, and the
complexity of the dataset. A balanced approach that incorporates elements from both
strategies can often lead to the best results.

5.6 CONCLUSION
The case study emphasizes the importance of comprehensive and diverse datasets.
The availability of a wide range of medical images representing different diseases
and conditions enables the system to generalize and adapt to new cases effectively.
Additionally, the inclusion of diverse patient populations helps to ensure that the
developed algorithms are applicable across various demographics, thus reducing
potential biases.
86 Data-Centric Artificial Intelligence for Multidisciplinary Applications

Furthermore, the case study highlights the need for robust and scalable infrastruc‑
ture to support the analysis and classification of medical images. This includes pow‑
erful computing resources, storage capacity, and secure data management systems.
Such infrastructure enables the efficient processing of large datasets and facilitates
collaboration among healthcare professionals and researchers [23].
In conclusion, the case study underscores the transformative potential of a
data‑centric approach in the analysis and classification of medical images. By lever‑
aging advanced algorithms, comprehensive datasets, and robust infrastructure,
healthcare professionals can enhance diagnostic accuracy, improve patient outcomes,
and accelerate medical research. The findings from this case study provide valuable
insights and serve as a foundation for further advancements in the field of medical
image analysis.

REFERENCES
[1] Roth, A., Wüstefeld, K. and Weichert, F., 2021. A data‑centric augmentation approach
for disturbed sensor image segmentation. Journal of Imaging, 7(10), p. 206.
[2] Bullock, J., Cuesta‑Lázaro, C. and Quera‑Bofarull, A., 2019. XNet: A convolutional
neural network (CNN) implementation for medical x‑ray image segmentation suitable
for small datasets. In: Medical Imaging 2019: Biomedical Applications in Molecular,
Structural, and Functional Imaging (Vol. 10953, pp. 453–463). SPIE, Bellingham, WA.
[3] Zahid, A., Poulsen, J.K., Sharma, R. and Wingreen, S.C., 2021. A systematic review
o emerging information technologies for sustainable data‑centric health‑care.
International Journal of Medical Informatics, 149, p. 104420.
[4] Kalyankar, P.A., Mulani, A.O., Thigale, S.P., Chavhan, P.G. and Jadhav, M.M., 2022.
Scalable face image retrieval using AESC technique. Journal of Algebraic Statistics,
13(3), pp. 173–176.
[5] Shamshad, F., Khan, S., Zamir, S.W., Khan, M.H., Hayat, M., Khan, F.S. and Fu,
H., 2023. Transformers in medical imaging: A survey. Medical Image Analysis, 88,
p. 102802.
[6] Wang, L., Xue, W., Li, Y., Luo, M., Huang, J., Cui, W. and Huang, C., 2017. Automatic
epileptic seizure detection in EEG signals using multi‑domain feature extraction and
nonlinear analysis. Entropy, 19(6), p. 222.
[7] Waugh, S.A., Purdie, C.A., Jordan, L.B., Vinnicombe, S., Lerski, R.A., Martin, P. and
Thompson, A.M., 2016. Magnetic resonance imaging texture analysis classification of
primary breast cancer. European Radiology, 26, pp. 322–330.
[8] Maharana, K., Mondal, S. and Nemade, B., 2022. A review: Data pre‑processing and
data augmentation techniques. Global Transitions Proceedings, 3(1), pp. 91–99.
[9] Preece, S.J., Goulermas, J.Y., Kenney, L.P. and Howard, D., 2008. A comparison of fea‑
ture extraction methods for the classification of dynamic activities from accelerometer
data. IEEE Transactions on Biomedical Engineering, 56(3), pp. 871–879.
[10] Aybike, U and Kilimci, Z.H., 2021. The prediction of chiral metamaterial resonance
using convolutional neural networks and conventional machine learning algorithms.
International Journal of Computational and Experimental Science and Engineering,
7(3), pp. 156–163.
[11] Dinh, A., Miertschin, S., Young, A. and Mohanty, S.D., 2019. A data‑driven approach
to predicting diabetes and cardiovascular disease with machine learning. BMC Medical
Informatics and Decision Making, 19(1), pp. 1–15.
Case Study Medical Images 87

[12] Bhattacharya, S., Maddikunta, P.K.R., Pham, Q.V., Gadekallu, T.R., Chowdhary, C.L.,
Alazab, M. and Piran, M.J., 2021. Deep learning and medical image processing for
coronavirus (COVID‑19) pandemic: A survey. Sustainable Cities and Society, 65,
p. 102589.
[13] Castaneda, C., Nalley, K., Mannion, C., Bhattacharyya, P., Blake, P., Pecora, A., Goy,
A. and Suh, K.S., 2015. Clinical decision support systems for improving diagnostic
accuracy and achieving precision medicine. Journal of Clinical Bioinformatics, 5(1),
pp. 1–16.
[14] Whang, S.E., Roh, Y., Song, H. and Lee, J.G., 2023. Data collection and quality chal‑
lenges in deep learning: A data‑centric ai perspective. The VLDB Journal, 32(4),
pp. 791–813.
[15] Hamid, O.H., 2023. Data‑centric and model‑centric AI: Twin drivers of compact and
robust Industry 4.0 solutions. Applied Sciences, 13(5), p. 2753.
[16] Panthong, R. and Srivihok, A., 2015. Wrapper feature subset selection for dimension
reduction based on ensemble learning algorithm. Procedia Computer Science, 72,
pp. 162–169.
[17] Taylor, K.I., Staunton, H., Lipsmeier, F., Nobbs, D. and Lindemann, M., 2020. Outcome
measures based on digital health technology sensor data: Data‑and patient‑centric
approaches. NPJ Digital Medicine, 3(1), p. 97.
[18] Taylor, K.I., Staunton, H., Lipsmeier, F., Nobbs, D. and Lindemann, M., 2020. Outcome
measures based on digital health technology sensor data: Data‑and patient‑centric
approaches. NPJ Digital Medicine, 3(1), p. 97.
[19] Kumar, A. and Gautam, S., 2022. Improving medical diagnostics with machine learn‑
ing: A study on data classification algorithms. International Journal of Advanced
Computer Research, 12(61), p. 31.
[20] Chavhan, P.G., Ratil, R.V. and Mahalle, P.N., 2023. An investigative approach of context
in internet of behaviours (IoB). In: International Conference on Emerging Trends in
Expert Applications & Security (pp. 333–343). Springer, Singapore.
[21] Daphal, P., Pokale, S., Chavhan, P., Wasatkar, N., Rathi, S., Dongre, Y. and Kolekar, V.,
2023. Human pose detection system using machine learning. International Journal of
Intelligent Systems and Applications in Engineering, 11(3), pp. 553–561.
[22] Kharate, N., Patil, S., Shelke, P., Shinde, G., Mahalle, P., Sable, N. and Chavhan,
P.G., 2023. Unveiling the resilience of image captioning models and the influence of
pre‑trained models on deep learning performance. International Journal of Intelligent
Systems and Applications in Engineering, 11(9s), pp. 1–7.
[23] Mahalle, P.N., Shinde, G.R., Ingle, Y.S. and Wasatkar, N.N. (2023). Data‑centric AI.
In: Data Centric Artificial Intelligence: A Beginner’s Guide: Data‑Intensive Research.
Springer, Singapore. doi:10.1007/978‑981‑99‑6353‑9_5
 6 Comparative Analysis
of Machine Learning
Classification
Techniques for Kidney
Disease Prediction
Jayashri Bagade, Nilesh P. Sable,
and Komal M. Birare

6.1 INTRODUCTION
The kidney’s primary job is to filter the body’s blood. Kidney disease is a silent
killer because kidney failure can occur without any warning signs or symptoms. The
definition of chronic renal disease is a deterioration in kidney function over months
or years. Diabetes and high blood pressure are common contributors to kidney dam‑
age. Globally, chronic kidney disease (CKD) is a serious health issue that affects
many people. People who can’t afford therapy for chronic renal disease may suffer
catastrophic effects if they don’t receive it. The most accurate test to assess kidney
function and the severity of CKD is the glomerular filtration rate (GFR). It can be
calculated using the blood creatinine level, as well as factors like age, gender, and
other details. Most of the time, becoming sick sooner is preferable. Consequently, it
is feasible to avoid major diseases. The kidney function of people with CKD gradu‑
ally deteriorates over time. It is a huge burden on the healthcare system due to its
rising frequency and high risk of developing end‑stage renal disease, which calls for
dialysis or kidney transplantation. A major worldwide health concern, CKD also has
a terrible prognosis for morbidity and mortality. However, CKD can be significantly
slowed down and serious complications can be avoided with early detection and
treatment. In order to effectively manage and treat the condition, it is imperative to
be aware of the signs and symptoms of renal disease. By leading a healthy lifestyle,
CKD can be prevented from progressing as quickly, by modifications like eating a
balanced diet, exercising frequently, abstaining from smoking and excessive alcohol
use, and managing underlying medical conditions like diabetes and hypertension.
Regular kidney function tests (KFTs), such as urine and blood tests, can also iden‑
tify CKD early on, allowing for quick management and intervention to stop further
kidney damage. The major objective of this research is to investigate datasets, flow

88 DOI: 10.1201/9781003461500-8
Comparative Analysis of Machine Learning Classification Techniques 89

FIGURE 6.1 Symptoms in CKD patients.

diagrams, and block diagrams to employ different algorithms to predict the develop‑
ment of renal illness.
Early identification of renal disease can help prevent irreparable kidney dam‑
age, albeit it may not always be possible. It’s critical to have a better understanding
of kidney disease symptoms to accomplish this goal. To predict the occurrence
of renal illness, the project involves analysing patient health and disease data,
comparing them using various indices and using machine learning classification
algorithms. Random Forests, K‑nearest Neighbour, Support Vector Machine, ADA
Boost, Gradient Boosting, Cat Boost, and Stochastic Gradient Boosting are some
of the classification methods used [1]. The data is categorized by the machine clas‑
sified into various classes, labels, and categories. Doctors often perform physical
examinations, evaluate the patient’s medical history, and then perform diagnostic
tests and treatments to identify the underlying cause of symptoms to diagnose an
illness.
With a fast‑rising patient population, CKD is currently the leading cause of death,
accounting for 1.7 million fatalities per year. Although there are many different diag‑
nostic techniques, this study uses machine learning because of its high accuracy.
Today, millions of people die from CKD, a condition that is quickly spreading and for
which there is now no timely, effective treatment. Patients with chronic renal disease
typically originate from middle‑ and low‑income nations.
Exercise, drinking water, and avoiding junk food are all advised. Figure 6.1
depicts the typical signs of chronic renal disease.

6.2 LITERATURE REVIEW

Due to its potential to enhance diagnostic precision, prognostic prediction, and thera‑
peutic outcomes, machine learning has attracted a lot of attention in recent years. For
effective treatment and management of kidney diseases, early detection and precise
90 Data-Centric Artificial Intelligence for Multidisciplinary Applications

disease progression prediction are essential. In 2015, Swathi Baby [2] suggested a
system based on predictive data mining for the development of an analysis and pre‑
diction tool for renal illness. Data on renal disease collected and analysed by Weka
and Orange software were used in the system. The study used a variety of machine
learning methods to anticipate and statistically analyse the likelihood of renal sick‑
ness, including Naive Bayes, J48, AD Trees, K Star, and Random Forest. Performance
indicators for each approach were computed and contrasted. The study’s findings
demonstrate that the K‑star and Random Forest (RF) algorithms outperformed the
opposition on the test dataset.
These methods have Receiver Operating Characteristic (ROC) values of 1 and 2,
respectively, and were discovered to produce models more quickly. A million people
perished in 2013 because of chronic renal disease. More people have CKD in devel‑
oping nations, where there are 387.5 million CKD patients overall, 177.4 million of
whom are men and 210.1 million of whom are women. These statistics demonstrate
that a significant portion of the population in emerging nations has chronic renal
disease and that this proportion is rising daily. For CKD to be treated at an early
stage, a lot of work has been put into early diagnosis. In this paper, we emphasize
accuracy while concentrating on machine learning prediction models for chronic
renal disease. When both kidneys are destroyed, a common type of kidney illness
known as CKD develops, and CKD patients must live with this condition for an
extended period. Any renal issue that could impair kidney function is referred to here
as kidney damage.
To forecast renal illnesses in 2015, Dayanand [3] suggested employing support
vector machines (SVMs) and artificial neural networks (ANNs). The primary objec‑
tive of this study was to compare the accuracy and execution times of these two
algorithms. According to the experimental data, ANN is performed better than SVM
in terms of accuracy. The Naive Bayes, SVM, and decision tree (DT) were used in
different machine learning algorithms for classification, while RF, logistic regres‑
sion, and linear regression were utilized in the medical sectors for regression and
prediction. Due to early‑stage diagnosis and prompt patient treatment, the death rate
can be reduced by the effective application of these algorithms. Patients with chronic
renal disease should maintain their clinical symptoms and engage in regular physi‑
cal activity. Researchers Ganapathi Raju, K Gayathri Praharshitha, and K Prasanna
Lakshmi completed a study in 2020 [4] that used different classification algorithms
on patient medical information to diagnose chronic kidney illness. The primary
objective of this study was to determine the classification algorithm that, based on
the classification report and performance indicators, would be most useful for diag‑
nosing CKD. In 2017, a team of researchers have used 14 attributes to predict CKD
and achieved 0.991 accuracy with a multiclass decision forest [8].
A team of researchers used a multiclass decision forest in 2017 to predict CKD
using 14 different attributes, and they were able to do so with an astounding accuracy
of 0.991 [6]. To increase accuracy, the researchers eliminated instances with missing
values and trained both a logistic regression model and a neural network. Overall
accuracy for these models was 0.975 and 0.960, respectively. Correlations between
the chosen attributes ranged from 0.2 to 0.8. From a medical standpoint, it’s critical
to consider the associations between characteristics and CKD. For instance, specific
Comparative Analysis of Machine Learning Classification Techniques 91

gravity has a correlation of 0.73 to the class and can both cause and be caused by
CKD. Eliminating these characteristics might result in a drop in accuracy. In 2017,
Sarica et al. [5] talked about the benefits of RF while considering any possible dis‑
advantages. More study on comparisons between this method and other widely used
classification systems is also encouraged, particularly in the early detection of the
transition from mild cognitive impairment (MCI) to Alzheimer’s disease (AD). This
study recommends a procedure that involves preprocessing of the data, collaborative
filtering for handling missing values, attribute selection, and CKD status prediction
using clinical data. The additional tree and RF classifiers displayed the highest accu‑
racy, obtaining 100% accuracy, with the least amount of bias towards the attributes.
In 2015, In 2021, Authors used the WEKA data mining tool to test eight machine
learning models [7]. The Naive Bayes, Multi‑layer Perception, and J48 algorithms
performed the best with accuracy scores of 0.95, 0.99, and 0.99, respectively, and
ROC scores of 1. The multilayer perceptron algorithm scored the highest in the study
using Kappa statistics, with a score of 0.99, followed by the decision table and J48
algorithms with scores of 0.97. El‑Houssainy et al. [8] examined the outcomes of vari‑
ous machine learning models and discovered that the Multiclass Decision Forest algo‑
rithm had the highest accuracy rate of about 99% for a condensed dataset with only 14
attributes. Supriya Aktar et al. [9] concentrated on using different machine learning
classification algorithms to increase the CKD diagnosis’ accuracy and shorten the
diagnosis process. The goal of the study was to categorize various stages of CKD
according to their severity. The performance of various algorithms, including radial
basis function (RBF,) RF, and Basic Propagation Neural Network, was examined by
the researchers. The analysis’s findings demonstrated that, with an accuracy of 85.3%,
the RBF algorithm performed better than the other classifiers. In a dataset of CKD,
Dilli Arasu and Thirumalaiselvi [10] have worked on missing values. The accuracy of
our model and the results of our predictions will both be lowered by missing values in
the dataset. They came up with a solution to the issue that by performing a recalcula‑
tion process on the CKD stages, they came up with unknown values.
They recalculated the values in place of the missing ones. With the aid of a machine
learning algorithm, Salekin and Stankovic [11] employ a novel approach to detect
CKD. They receive results based on a dataset with 400 records and 25 attributes that
indicate whether a patient has CKD or not. To obtain results, they employ neural
networks, RFs, and k‑nearest neighbours. They use the wrapper method for feature
reduction. For feature reduction, they use the wrapper method which detects CKD
with high accuracy. The effects of class imbalance during data training for the devel‑
opment of neural network algorithms for the treatment of chronic renal disease are
examined by Yildirim [12]. In the suggested work, comparative research was carried
out using a sampling algorithm. This study demonstrates how sampling algorithms
can help classification algorithms perform better. It also shows that the learning rate
is a crucial factor that has a big influence on multilayer perceptron. Sharma et al. [13]
tested 12 different classification algorithms on a dataset containing 400 records and
24 attributes. They have compared their estimated findings with the actual results to
ascertain the precision of their forecasts. They used evaluation standards like preci‑
sion, sensitivity, accuracy, and specificity. The DT approach offers a precision of 1,
specificity of 1, and sensitivity of 0.9720 with an accuracy of up to 98.6%.
92 Data-Centric Artificial Intelligence for Multidisciplinary Applications

6.3 METHODS FOR CLASSIFICATION

In this study, the presence of chronic renal illness in people was predicted using
several classification methods. The classifiers employed for classifications are SVM,
K‑Nearest Neighbour (KNN), ADA Boost (ADA), Cat Boost CAT, DT, and (RF.
Each classifier projected the chance of the disease using the chronic kidney/renal
disease dataset. After being evaluated for accuracy, precision, and F‑measure, the
classifier with the best performance was picked. The recommended approach makes
use of a predictive data mining framework. The workflow is depicted in Figure 6.2.
Many hospitals and medical labs throughout the city provided the data for this
investigation. The synthetic KFT dataset, which may be used to study various dis‑
eases, was made using these data. The dataset used in the study, which examined 401
cases, contained six variables including age, gender, urea, creatinine, and GFR. Data
on renal diseases and how they impact kidney function are included in the KFT data‑
set. A dataset of patient medical details was used to test different classification tech‑
niques, and the effectiveness of each algorithm was compared. This analysis served
as the foundation for the prediction. The model was trained and examined using a
training dataset with 320 rows. The model’s performance on an unknown dataset was
assessed using the testing dataset, which had 80 rows. To predict the likelihood of the
disease, the dataset for CKD was gathered and applied to each classifier. Based on
the accuracy rate, precision, and F‑measure values gleaned from the entire analysis,
each classifier’s performance was assessed. Quite a few algorithms, including SVM
and KNN, are implemented in the project’s architecture.
The choice to use an SVM was made based on the problem’s nature, which
may not be linearly separable. The SVM algorithm with a nonlinear kernel is a
good choice in these circumstances. On the other hand, KNN is renowned for its

FIGURE 6.2 Workflow of experimentation.

Comparative Analysis of Machine Learning Classification Techniques 93

FIGURE 6.3 Algorithmic approach.

resistance to noise in training data, though its effectiveness might depend on the
number of training examples used. The accuracy of the KNN algorithm depends
on choosing the best value for the K parameter, which establishes the number of
nearest neighbours and the distance metric to be applied. Calculations in machine
learning can take a while, particularly when determining how far apart each
instance is from all training instances. To handle categorical features of a dataset,
RF, which is essentially a collection of DTs combined, can be used. Many train‑
ing examples and high‑dimensional spaces can both be handled by this algorithm.
Certain criteria must be established to use RF. The algorithmic approach is shown
in Figure 6.3.

6.3.1 Support Vector Machines (SVM)

The powerful, cutting‑edge regression method SVM can be used with both linear
and nonlinear data. SVM is utilized for binary and multiclass classification in Oracle
Data Mining. The main benefit of SVM is its capacity to infer, implicitly, in a modi‑
fied (nonlinear) feature space the distance between a molecule and the hyperplane
without changing the original descriptors. This is feasible as a result of the “kernel
trick.” The most popular Gaussian kernel, also referred to as the radial basis function
94 Data-Centric Artificial Intelligence for Multidisciplinary Applications

kernel, was used in this study. The SVM algorithm’s objective is to establish the
optimal decision boundary or line that can divide n‑dimensional space into classes
so that we may quickly place.

6.3.2 K‑Nearest Neighbour (KNN)

The K‑Nearest Neighbour algorithm (K‑NN) is a nonparametric pattern recognition
model for classification and regression. It operates by finding the K closest training
instances in the feature space, regardless of whether the input is for classification or
regression. KNN is an illustration of instance‑based learning. The algorithm out‑
puts a class membership for KNN classification, where a 75% vote of neighbours
is necessary for classification, when K=3, the class contains only the three near‑
est neighbours. Each neighbour is given a weight of 1/d in a traditional weighting
system, where d is the distance from the data points. The Euclidean distance, or
the shortest distance in a straight line between two neighbours, is calculated by the
K‑NN algorithm. K‑NN’s lack of sensitivity to the data’s local configuration values is
a disadvantage, though. Before using the K‑NN method in the feature space, feature
extraction is done to turn raw data into a set of features.
The k‑nearest neighbours’ algorithm, sometimes referred to as KNN or K‑NN, is
a supervised learning classifier that employs proximity to produce classifications or
predictions about the grouping of a single data point. Although it can be applied to
classification or regression issues, it is commonly employed as a classification algo‑
rithm because it relies on the idea that comparable points can be discovered close to
one another.

6.3.3 ADA Boost Technique

Popular and effective, ADA has been used extensively in a variety of industries,
including computer vision, natural language processing, and bioinformatics. The
algorithm has become a useful tool for resolving challenging issues because of its
propensity to adapt to changing data and enhance performance over time. Its use
in ensemble methods has also improved the robustness and accuracy of machine
learning models. Many researchers and practitioners in the field rely on the ADA
algorithm because of its adaptability and efficiency.

6.3.4 Cat Boost Technique

A brand‑new open‑source machine learning algorithm called Cat Boost was created
by Yandex in 2017. This algorithm, which is based on the idea of trees and gradient
boosting, is particularly helpful for categorical boosting, application ranking, and
various recommendation systems. The Cat Boost algorithm is well‑known for being
simple to implement and simple to use, and it is frequently used to solve classifica‑
tion and regression problems. This algorithm also has the benefit of being effective
with small and heterogeneous datasets. The Cat Boost algorithm’s capacity to handle
categorical features automatically and effectively is perhaps its most noteworthy ben‑
efit, making it a useful tool for data analysis in a variety of contexts. Both categorical
Comparative Analysis of Machine Learning Classification Techniques 95

and numerical features can be handled by the gradient boosting variation known as
Cat Boost. To transform categorical data into numerical features, there is no need for
feature encoding methods like One‑Hot Encoder or Label Encoder.

6.3.5 Decision Tree
One common supervised learning method for both classification and regression tasks
is the DT algorithm. Although it can be used to solve regression issues, classification
tasks are where it is most frequently applied. The algorithm uses a classifier struc‑
ture resembling a tree, with internal nodes denoting dataset features and internal
branches denoting decision rules. DTs use an inquiry‑based approach, in which ques‑
tions are posed to ascertain whether or not a specific attribute is present. The data
are then divided into subtrees and further examined for classification or regression
tasks using the results.

6.3.6 Random Forest
The popular machine learning algorithm RF is a part of the supervised learning
methodology. It can be applied to ML issues involving both classification and regres‑
sion. It is predicated on the idea of ensemble learning, which is the act of integrating
various classifiers to address a complicated issue and enhance the model’s perfor‑
mance. According to what its name implies, “Random Forest is a classifier that con‑
tains a number of DTs on various subsets of the given dataset and takes the average
to improve the predictive accuracy of that dataset.” Instead of depending on a single
DT, the RF uses forecasts from each tree and predicts the result based on the votes
of the majority of predictions.
Higher accuracy and overfitting are prevented by the larger number of trees in
the forest. Some DTs may predict the correct output, while others may not because
the RF combines numerous trees to forecast the class of the dataset. But when
all the trees are combined, they forecast the right result. For the dataset’s feature
variable to predict true outcomes rather than a speculated result, there should be
some actual values in the dataset. Each tree’s predictions must have extremely low
correlations.

6.4 RESULT
The effectiveness of each algorithm was assessed after a study was conducted using
a variety of algorithms. The accuracy values for each algorithm are displayed on
a graph that was made to help convey this information in an understandable and
concise manner. The graph gives a visual representation of how each algorithm per‑
forms, making it simple to compare them and determine which algorithm is the most
accurate for the given dataset.
Data analysis frequently employs this method of presenting findings through
graphs because it makes complex information easier to understand. It is simpler to
spot patterns and trends. Graphs can also aid in the simplification of complex infor‑
mation to make it more understandable to a wider audience.
96 Data-Centric Artificial Intelligence for Multidisciplinary Applications

FIGURE 6.4 Comparison of accuracy of models.

TABLE 6.1
Performance Comparison of Models
Accuracy
Sr. No Classification Model Training Testing Precision Recall F1 Score
1 KNN 0.77 0.67 0.74 0.71 0.72
2 SVM 0.94 0.95 0.99 0.93 0.96
3 Random forest 1.0 0.975 0.96 0.94 0.98
4 Decision tree 0.97 0.94 0.93 0.97 0.95
5 ADA Boost 1.0 0.97 0.96 1.00 0.98
6 Cat Boost 1.0 0.966 0.96 0.99 0.97

In this situation, the graph showing the accuracy values for each algorithm can
assist academics and medical professionals in deciding which algorithm is best for
identifying a specific disease or condition. The development of more precise diag‑
nostic tools or an improvement in patient care for the specific condition can both
benefit from this information. Overall, the effectiveness of data analysis in healthcare
and other fields can be greatly improved using graphs and other visual aids when data
is presented visually than when it is presented in its raw form.
The machine learning models are trained on the KFT dataset to validate the per‑
formance of the model. The models trained on 320 data records are used to test 80
unseen data records. Performance of the models is evaluated using four evaluation
parameters namely accuracy, precision, recall, and F1 score. The results are presented
in Table 6.1. A comparison of the accuracy of the model is shown in Figure 6.4.

6.5 CONCLUSIONS AND FUTURE SCOPE

In the medical field, data mining and machine learning have found uses, including
in the prognosis of CKD. To predict the likelihood of CKD, this study introduces a
Comparative Analysis of Machine Learning Classification Techniques 97

novel decision support system that makes use of machine learning classifiers. The
classifiers were successful in predicting other diseases in addition to CKD. The
effectiveness of three different classifiers used in the study to predict the presence of
CKD was compared. The findings demonstrated that in terms of predicting CKD, the
ADA Boost and RF classifiers performed better than the SVM, DT, KNN, and CAT
Boost classifiers. Overall, the study has overcome earlier shortcomings and increased
the precision of CKD prediction. The study’s results are encouraging and serve as a
foundation for additional study in this area.
The current techniques for anticipating chronic renal disease are deemed suf‑
ficient, despite some drawbacks. However, as the table below illustrates, research
is still being done to increase the precision of kidney disorder prediction and iden‑
tification. A more advanced CKD prediction system is still needed despite current
initiatives. Given that this area of research is still largely unexplored, a decision sup‑
port system that can aid in the early detection of chronic renal disease is especially
necessary.
This study uses a variety of classification methods to identify CKD including
RFs, KNN, SVM, ADA Boost, DT, and Cat Boost. These classifiers’ performance
can be evaluated and contrasted with that of other classifiers. For prompt treatment
to begin and to stop the disease from worsening, early detection of CKD is essential.
Therefore, early disease detection and prompt treatment are crucial for the medical
industry. Alternative classifiers can be investigated and assessed in upcoming studies
to find better approaches to objective function problems.

REFERENCES
[1] Konstantina Kourou, Themis P. Exarchosa, Konstantinos P. Exarchos, Michalis V.
Karamouzis, Dimitrios I. Fotiadis: “Machine learning applications in cancer progno‑
sis and prediction”, Computational and Structural Biotechnology Journal, Vol. 13,
pp. 8–17 (2015).
[2] P. Swathi Baby, T. Panduranga Vital: “Statistical analysis and predicting kidney dis‑
eases using machine learning algorithms” International Journal of Engineering
Research and Technology, Vol. 4, (2015).
[3] S. Vijayarani, S. Dhayanand: “Kidney disease prediction using SVM and ANN algo‑
rithms”, International Journal of Computing and Business Research, Vol. 6, (2015).
[4] Ganapathi Raju, K. Gayathri Praharshitha, K. Prasanna Lakshmi: “Prediction of
chronic kidney disease (CKD) using data science”, International Conference on
Intelligent Computing and Control Systems (ICCS) (2019).
[5] Alessia Sarica, Antonio Cerasa, Aldo Quattrone: “Random forest algorithm for the
classification of neuroimaging data in alzheimer’s disease”, Front Aging Neurosci. 2017
Oct 6; 9: 329. doi: 10.3389/fnagi.2017.00329. PMID: 29056906; PMCID: PMC5635046
(2017).
[6] N.V. Ganapathi Raju, K. Prasanna Lakshmi, K. Gayathri Praharshitha: “Prediction
of chronic kidney disease (CKD) Using data science”, International Conference on
Intelligent Computing and Control Systems (ICCS) (2017).
[7] Gazi Mohammed Ifraz, Muhammad Hasnath Rashid, Tahia Tazin, Sami Bourouis,
Mohammad Monirujjaman Khan: “Comparative analysis for prediction of kidney dis‑
ease using intelligent machine learning methods” Comput Math Methods Med. 2021
Dec 3; 2021: 6141470. doi: 10.1155/2021/6141470. Retraction in: Comput Math Methods
Med. 2023 Nov 1; 2023: 9864519. PMID: 34899968; PMCID: PMC8664508, (2021).
98 Data-Centric Artificial Intelligence for Multidisciplinary Applications

[8] El‑Houssainy A. Ready, Ayman S. Anwar: “Prediction of kidney disease stages using
data mining”. Informatics in Medicine Unlocked, Vol. 15 (2019).
[9] Suraiya Aktar, Abhijit Pathak, Abrar Hossain Tasin: “Chronic kidney disease (CKD)
Prediction using data mining techniques”, In book: Advances in Intelligent Systems and
Computing (pp. 976–988). Publisher: Springer, Cham. (2021).
[10] S Dilli Arasu, R Thirumalaiselvi: “Review of chronic kidney disease based on data
mining techniques”, International Journal of Applied Engineering Research ISSN
0973‑4562 Volume 12, Number 23 pp. 13498–13505 (2017).
[11] Asif Salekin, John Stankovic: “Detection of Chronic Kidney Disease and Selecting
Important Predictive Attributes”, 2016 IEEE International Conference on Healthcare
Informatics (ICHI), (2016).
[12] Elias Dritsas, Maria Trigka: “Machine learning techniques for chronic kidney disease
risk prediction.” Pinar Yildirim: “Chronic Kidney Disease Prediction on Imbalanced
Data by Multilayer Perceptron: Chronic Kidney Disease Prediction”, IEEE 41st Annual
Computer Software and Applications Conference (COMPSAC), (2017).
[13] Sahil Sharma, Vinod Sharma, Atul Sharma: “Performance Based Evaluation of Various
Machine Learning Classification Techniques for Chronic Kidney Disease Diagnosis”,
International Journal of Modern Computer Science (IJMCS), Vol. 4, (2016).
 7 Fusion of Multi‑Modal
Lumber Spine Scans
Using Convolutional
Neural Networks
Bhakti Palkar

7.1 INTRODUCTION: BACKGROUND AND DRIVING FORCES

The spine supports all the activities like lifting, walking, sitting and standing.
Number of spine patients is rising across the globe. In India, more than 200,000
spine patients are treated in a year. More than 30,000 cases are of only spinal cord
injury due to accidents [1]. Earlier, back pain used to be a problem for only elderly
people, but nowadays, we observe back pain in children and young adults also. It is
mainly because of an unhealthy lifestyle. Earlier kids and young adults used to play
sports outside, but now they are glued to television or mobile phones and frequently
in an abnormal posture on a couch and consuming a lot of junk food. This causes
lower back problems. Back muscles support the spine and its strength is important to
the human body. When we lift something heavy, major stress is handled by the back
muscles. If the back muscles are weak, then the stress is transferred to the spine ver‑
tebras and discs. This leads to various spine‑related diseases like degenerative spine
diseases, spine Spondylosis, Sciatica and scoliosis. World Spine Day is observed on
16 October every year. In 2017, a major metropolitan four‑city survey was conducted
in India to identify categories and types of spine patients. Nearly 75% of patients of
the surveyed people suffered from lower back pain and the age group of these people
was 16–34 [2]. The patients are increasing rapidly and so is the quantity of medical
scans. The rapid growth in the quantity of medical images being captured world‑
wide is also because companies like GE Healthcare, Siemens and Philips are coming
up with technically advanced, easy and user‑friendly devices. Magnetic resonance
imaging (MRI), X‑ray, computed tomography (CT), ultrasound, mammograms,
positron emission tomography (PET), Single‑photon emission computed tomogra‑
phy (SPECT) PET, SPECT are some common medical imaging techniques. Each
technique has its own traits. Depending on the body organ being observed, imaging
modality is prescribed by a doctor. In some situations where one imaging modality is
not enough to arrive at some conclusion, doctors suggest multiple modality images.
The pervasiveness of medical images influences the necessity of accurate analysis

DOI: 10.1201/9781003461500-9 99
100 Data-Centric Artificial Intelligence for Multidisciplinary Applications

FIGURE 7.1 Quantity of papers available in PubMed database on “medical image fusion”.
Source: PubMed Dataset.

of these images. The research area “medical image fusion” is growing vastly [3]. In
Figure 7.1, we can observe the increasing number of publications based on “Medical
Image Fusion” in PubMed database from year the 2000 to June 2023.
Key Contribution: In this research work, lumbar spine diseases of 12 different
types are considered for fusion. A novel deep learning and wavelet‑based approach
is used for CT and MRI fusion of spine images to generate one image that contains
details from both MR and CT images.

7.2 RELATED WORK

Image fusion is a very broad term. It is applied on any two images to integrate their
contents into one image. All the methods of fusion are generally sorted into four
categories: (1) Spatial domain; (2) Pyramid based; (3) Frequency domain transforms;
(4) Neural network.
A variety of spatial domain techniques like simple average, minimum selec‑
tion, maximum selection, Max‑Min, simple block replace, weighted averaging,
intensity‑hue‑saturation, Brovey, principal component analysis (PCA) and guided
filtering are used for fusion. Being simple and easy to implement is the main distin‑
guishing factor of these methods. However, these methods may show blurred output,
sometimes making them not suitable for medical applications. In this category, only
intensity‑hue‑saturation combined with PCA is applied on PET and MR images of
brain to detect Alzheimer’s [4].
Pyramid‑based fusion techniques [5–16] generate images of decent quality. These
techniques have been applied mostly on general images. The output depends on the
number of decomposition levels. The higher the number of decomposition levels,
the better would be the output here. Only a morphological pyramid in this cate‑
gory is used on CT and MR images of the human brain for tumour diagnosis [16].
Sometimes, pyramid‑based fusion methods show blocky effect on output images. To
deal with this problem, the Laplacian pyramid is combined with a Discrete Cosine
Transform, and then, it is applied on MRIs and CT scans of the brain [17].
Fusion of Multi-Modal Lumber Spine Scans 101

The third category of fusion methods is “frequency domain‑Transform based

fusion”. The most extensively used transform in the medical domain is discrete
wavelet transform (DWT). It is typically applied on CT and MR images of the brain
[18–25]. It is also combined with PCA as a fusion rule to detect brain cancer cells
using CT and MR images [26]. We can generate wavelets using other transforms and
use them for fusion in place of DWT [27,28]. In spite of providing a better signal‑
to‑noise ratio, DWT is sometimes not preferred by researchers because it lacks
phase information and it is a shift variant. Complex wavelet transform is also quite
popular among researchers. The dual tree complex wavelet transform (DT‑CWT)
overcomes the drawbacks of DWT but it requires more processing power, and more
memory than DWT because of the dual trees [29]. DT‑CWT is combined with
particle swarm optimization for brain diagnosis using CT and MR images [30,31].
Rotated wavelet transform (RWT) is slightly different from DWT. It preserves tex‑
ture and displays high contrast. However, edge orientations are limited in the output
of RWT. It is also applied to CT‑MR images of the brain [32]. Stationary wavelet
transform (SWT) removes “shift variance” drawback of DWT like DT‑CWT. But
its computational complexity is very high and it needs larger storage space. SWT
has been used to fuse CT and MR images of the brain [33]. Lifting wavelet trans‑
form reduces the memory requirement of DWT but it is very time‑consuming. It
has been tested on CT‑MRI, MRI‑PET and CT‑SPECT scans of brain [34]. All
the variants of wavelet transform cannot extract curves very efficiently. Curvelet
transform efficiently represents curves of the edges with a lesser number of coef‑
ficients. However, it generates redundant features and requires a much longer time
to calculate coefficients. Curvelet transform has been applied to CT‑MR images
of the brain, abdomen and CT‑mammograms of the breast [35]. Contourlet trans‑
form is good at capturing geometric structures in images. It lacks shift‑invariance,
results in ringing artefact and cannot deal with mis‑registration which in turn results
in image distortion. Contourlet transform has been applied to CT‑MR images of
lungs, T1MRI‑T2MRI of knees and X‑Ray‑CT of eyes [36,37]. Non‑subsampled
contourlet transform (NSCT) deals with the problems of “mis‑registration”, but it
is very time‑consuming. NSCT has been applied to MRI‑PET images of the brain
[38]. The fourth category of fusion methods is “neural network based fusion” [39].
Artificial neural network is mostly clubbed with other methods like wavelet and
fuzzy logic for fusion [40]. Pulse‑coupled neural network (PCNN) does not require
any training and it produces a minimal number of features even for large images.
It faces difficulty in estimating the optimum values for its large number of param‑
eters. It is applied to CT‑MR images of the brain [41]. PCNN is combined with
non‑subsampled shearlet transform [42], ripplet transform [43] and also with NCST
[44] for CT and T2–MR fusion. In the past 4 years, people have been using deep
networks to extract features [45,46]. Finding a good fusion algorithm in the medical
domain is still a challenge because a fusion method that works for one combination
of multi‑modal images may fail completely for other modalities. It is also observed
that “brain” is the most researched organ in this research area. Almost little‑to‑no
attention is paid to “spine” which is also a crucial body part.
102 Data-Centric Artificial Intelligence for Multidisciplinary Applications

7.3 PROPOSED WORK

This section introduces a new method to fuse multi‑modal images using two very
much known deep learning frameworks – VGG19 and AlexNet. These deep frame‑
works are used in combination with 2D‑DWT.

7.3.1 2D‑Discrete Wavelet Transform

In the first step, the Image goes through the low pass filter and the high pass filter
separately. The outputs are then downsampled column‑wise thereby reducing the
number of columns by half. The downsampled images are then sent through the low
pass filter and the high pass filter separately, and then again, the outputs are downs‑
ampled row‑wise. This process generates four different images which are one‑fourth
of the original image in size. One image shows approximation coefficients and the
other three images show horizontal, vertical and diagonal features. Figure 7.2 shows
the result of applying 2D‑DWT on the CT image.

FIGURE 7.2 2D‑ DWT of CT image of the lumber spine.

Fusion of Multi-Modal Lumber Spine Scans 103

7.3.2 Visual Geometry Group (VGG)‑19 Network

The visual geometry group (VGG) network came out in 2014 as the second winner
of a competition. Vgg19 is one of the variants of VGG. The number 19 is the number
of convolution layers used in the network. Each convolution layer has a RELU acti‑
vation function with it. The layers are divided into five sets. Very tiny filters of size
3 × 3 are used in each layer. The size of the convolution layer is doubled with each
passing set. After each set of convolution layer, a 2 × 2 max pool layer is used with
stride 2. It reduces the image size by half. After the last set of convolution layer, the
image size reduces to 7 × 7. The image is then flattened and three fully connected
layers are used.
We can use pre trained VGG‑19 network trained on billions of images of ImageNet
dataset [47]. It can detect 1000 different classes. All the images are of size 224 × 224.
Since the network is trained over a variety of images, it has learned very rich features
from the images. That is the main reason it is used in the transfer learning approach.
Figure 7.3 shows VGG19 architecture.

7.3.3 AlexNet Network
AlexNet is also a convolutional neural network like VGG[48]. Alexnet is the winner
of the ILSVRC competition held in 2012. Figure 7.4 shows AlexNet architecture.
It is much smaller in size than VGG‑19 with just five convolutional layers and
three fully connected layers. Each Convolutional layer has a RELU activation func‑
tion with it except the last layer. Max‑pooling layer is used after the first, second
and fifth convolutional layers which are of size 3 × 3 with stride 2. It has three fully
connected layers at the end. Like VGG19, AlexNet is also trained over the ImageNet
dataset.

7.3.4 Wavelet+Deep Fusion

Figure 7.5 shows the proposed “Wavelet+Deep Fusion” Architecture.
CT and MRI images of the lumber spine are first aligned with each other and then
they are used for fusion. Both the images are decomposed into four different images ‑
approximation coefficient (LL), horizontal coefficient (LH), vertical coefficient (LV)
and diagonal coefficient (LD) using 2D‑DWT. A convolutional neural network is
used to fuse similar coefficients of the two images. So, “Deep Fusion” is invoked
four times to fuse the four different coefficients. Once the four different fused images
are obtained, inverse DWT is used on them to get the final fused image. Figure 7.6
demonstrates how “Deep Fusion” is used to fuse two images.

7.4 RESULT AND DISCUSSION

All techniques experimented with in the research work have used the SpineWeb data
set [49]. It contains CT and MR images of the lumber spine of 20 patients. The data‑
set covers various degenerative lumber spine diseases. Each CT image is registered
(aligned) with an MR image of the same pair using a landmark‑based registration
104 Data-Centric Artificial Intelligence for Multidisciplinary Applications

FIGURE 7.3 VGG19 architecture.

Fusion of Multi-Modal Lumber Spine Scans 105

FIGURE 7.4 AlexNet architecture.

FIGURE 7.5 Wavelet+deep CNN architecture.

106 Data-Centric Artificial Intelligence for Multidisciplinary Applications

FIGURE 7.6 Deep fusion.

technique. Lumbar spine fusion using “wavelet with VGG‑19” and “Wavelet with
Alexnet” have been compared with the following conventional methods.

1. Spatial Domain Technique‑Simple Average

2. Discrete Cosine Transform (DCT)
3. Hybrid Wavelet Transform (HWT)
4. Stationery Wavelet Transform (SWT)
5. Dual Tree‑Complex Wavelet Transform (DT‑CWT)
6. Discrete Cosine Transform‑Laplacian Pyramid (DCT‑LP)

7.4.1 Fused Images

Patient 1: Figure 7.7 shows how the CT image is aligned with the MR image for
Patient 1, and Figure 7.8 shows fused images generated using the CT and MRI of
Patient 1.
Patient 2: Figure 7.9 shows how the CT image is aligned with the MR image for
Patient 2, and Figure 7.10 shows fused images generated using CT and MRI of the
same patient.

7.4.2 Quantitative Analysis of All Techniques

The quality of all fused images is judged using three performance metrics: entropy,
spatial frequency and standard deviation [50]. Entropy indicates information con‑
tent. Tables 7.1–7.3 show values of entropy, standard deviation and spatial frequency,
respectively for five patients using all the methods.
The high value of entropy indicates that the image has a good amount of informa‑
tion. High value of standard deviation indicates that the image has very good contrast.
Fusion of Multi-Modal Lumber Spine Scans 107

FIGURE 7.7 (a) Moving image (CT)‑control point selection, (b) fixed image (MR)‑control
point selection, and (c) CT registered with MR image.

The high value of spatial frequency indicates good quality. “Wavelet+Vgg19” tech‑
nique showed the best values (shown in bold in every row) for all the patients.

7.5 CONCLUSION
Lumber Spine CT and MRI scans are merged into one image to observe L1, L2, L3,
L4 and L5 vertebras, discs, spinal cord, nerves and tissues in one image. A novel
technique based on wavelet and CNN is introduced in this research work. VGG19
108 Data-Centric Artificial Intelligence for Multidisciplinary Applications

FIGURE 7.8 Patient 1 fused images: (a) simple average, (b) DCT, (c) hybrid wavelet trans‑
form, (d) wavelet+alexnet, (e) SWT, (f) DCT‑LP, (g) DT‑CWT, and (h) Wavelet+VGG19.

and Alexnet are combined with 2D‑DWT to generate a fused image. Extensive
experimentation is done to observe difference in fused images generated using eight
different techniques, of which six are conventional techniques and two are novel
techniques. Registration of CT with MRI is done using a landmark‑based registration
technique, and then it is fused with MRI. Three evaluation performance metrics –
entropy, spatial frequency and standard deviation are used to compare all the meth‑
ods. It is observed that fused images generated using Wavelet+VGG19 technique
showed the highest values for all three parameters which indicate that this fused
image has at most information content, better contrast and quality than all the other
images. Medical experts can look at this image only instead of looking at CT and
MRI images of the spine. VGG19 can be replaced by other convolutional neural
networks like RESNET and GoogleNet to observe the difference in fused images.
Fusion of Multi-Modal Lumber Spine Scans 109

FIGURE 7.9 (a) CT, (b) MRI and (c) CT registered with MR image.

FIGURE 7.10 Patient 2 fused images. (a) Simple average, (b) DCT, (c) hybrid wavelet trans‑
form, (d) wavelet+alexnet, (e) SWT, (f) DCT‑LP, (g) DT‑CWT, and (h) wavelet+VGG19.
110 Data-Centric Artificial Intelligence for Multidisciplinary Applications

TABLE 7.1
Entropy
Patient Simple
No. Average DCT HWT Wavelet+AlexNet SWT DCT‑LP DT‑CWT Wavelet+VGG19
1 7.30 7.30 7.30 7.31 7.52 7.34 7.28 7.74
2 7.56 7.56 7.56 7.58 7.78 7.60 7.55 7.85
3 7.27 7.27 7.27 7.28 7.44 7.29 7.29 7.80
4 7.55 7.55 7.55 7.55 7.41 7.57 7.53 7.72
5 7.06 7.06 7.06 7.08 7.37 7.09 7.07 7.43

TABLE 7.2
Standard Deviation
Patient Simple
No. Average DCT HWT Wavelet+AlexNet SWT DCT‑LP DT‑CWT Wavelet+VGG19
1 51.16 50.52 51.16 48.42 62.03 49.58 53.04 63.53
2 47.00 63.07 47.07 47.57 62.09 48.28 46.81 64.29
3 42.30 46.39 42.35 42.35 54.12 42.38 42.56 60.63
4 58.80 59.82 58.87 58.88 72.47 59.12 59.05 73.23
5 35.59 46.36 35.66 36.01 47.08 36.28 35.53 48.99

TABLE 7.3
Spatial Frequency
Patient Simple
No. Average DCT HWT Wavelet+AlexNet SWT DCT‑LP DT‑CWT Wavelet+VGG19
1 7.00 7.08 7.00 5.24 5.88 5.55 3.56 9.96
2 13.42 13.45 13.44 11.15 10.86 12.31 6.49 14.35
3 5.23 5.34 5.24 2.94 4.86 3.10 2.87 9.71
4 9.07 9.09 9.08 6.03 7.64 6.63 4.85 11.62
5 6.98 7.01 7.00 6.10 5.69 6.56 3.39 7.53

REFERENCES
[1] http://timesofindia.indiatimes.com /articleshow/64992298.cms?utm_source=
contentofinterest&utm_medium=text&utm_campaign=cppst
[2] https://timesofindia.indiatimes.com/city/mumbai/73‑patients‑with‑spine‑problems‑
have‑lower‑back‑complaint‑survey/articleshow/61137025.cms
[3] https://pubmed.ncbi.nlm.nih.gov/?term=medical+image+fusion
[4] Changtao He, Quanxi Liu, Hongliang Li, Haixu Wang, Multimodal medical image
fusion based on IHS and PCA, Procedia Engineering, 7, 2010, pp. 280–285.
Fusion of Multi-Modal Lumber Spine Scans 111

[5] Mahesh Malviya, Sanju Kumari, Srikant Lade, Image fusion techniques based on pyra‑
mid decomposition, International Journal of Artificial Intelligence and Mechatronics,
4, 2014, pp. 127–130.
[6] P. Burt, E. Adelson, Laplacian pyramid as a compact image code, IEEE Transactions
on Communications, 31, 1983, pp. 532–5407.
[7] Jianguo Sun, Qilong Han, Liang Kou, Liguo Zhang, Kejia Zhang, and Zilong Jin,
Multi‑focus image fusion algorithm based on Laplacian pyramids, Journal of the
Optical Society of America 35, 2018, pp. 480–490.
[8] H. Olkkonen, P. Pesola, Gaussian pyramid wavelet transform for multiresolution analy‑
sis of images, Graphical Models and Image Processing, 58, 1996, pp. 394–398.
[9] P. Burt, A gradient pyramid basis for pattern selective image fusion, The Society for
Information Displays (SID) International Symposium Digest of Technical Papers, 23,
1992, pp. 467–470.
[10] A. Toet, Image fusion by a ratio of low‑pass pyramid, Pattern Recognition Letters, 9,
1996, pp. 245–253.
[11] H. Anderson, A filter‑subtract‑decimate hierarchical pyramid signal analyzing and syn‑
thesizing technique, U.S. Patent 718–104, 1987.
[12] L. C. Ramac, M. K. Uner, P. K. Varshney, Morphological filters and wavelet based image
fusion for concealed weapon detection, Proceedings of SPIE, 3376, 1998, pp. 110–119.
[13] M. D. Jasiunas, D. A. Kearney, J. Hopf, Image fusion for uninhabited airborne vehicles,
In: Proceedings of IEEE International Conference on Field Programmable Technology,
2002, pp. 348–351.
[14] S. Marshall, G. Matsopoulos, Morphological data fusion in medical imaging, In: IEEE
Winter Workshop on Nonlinear Digital Signal Processing, IEEE, 1993, pp. 6–1.
[15] K. Mikoajczyk, J. Owczarczyk, W. Recko, A test‑bed for computer‑assisted fusion of
multi‑modality medical images, In: Chetverikov, D., Kropatsch, W.G. (eds) Computer
Analysis of Images and Patterns. CAIP 1993. Lecture Notes in Computer Science, vol
719, Springer, 1993, pp. 664–668.
[16] G. Matsopoulos, S. Marshall, J. Brunt, Multiresolution morphological fusion of MR
and CT images of the human brain, In: Vision, Image and Signal Processing, IEE
Proceedings, vol. 141, IET, 1994, pp. 137–142.
[17] V. P. S. Naidu, Bindu Elias, A novel image fusion technique using DCT based Laplacian
pyramid, International Journal of Inventive Engineering and Sciences (IJIES), 1, 2,
2013, pp. 1–18.
[18] S. G. Mallat, A theory for multiresolution signal decomposition, the wavelet represen‑
tation, IEEE Transaction on Pattern Analysis and Machine Intelligence, 11, 7, 1989,
pp. 674–693.
[19] Yong Yang, Dong Sun Park, Shuying Huang, Zhijun Fang, Zhengyou Wang, Wavelet
Based Approach for Fusing Computed Tomography and Magnetic Resonance Images,
IEEE, 2009, pp. 5770–5772.
[20] Yong Yang, Dong Sun Park, Shuying Huang, Nini Rao, Medical image fusion via
an effective wavelet based approach, Journal on Advances in Signal Processing,
579341 2010, 44.
[21] Guihong Qu, Dali Zhang, Pingfan Yan, Medical image fusion using two‑dimensional
discrete wavelet transform, In: Proc. SPIE 4556, Data Mining and Applications, 2001.
doi:10.1117/12.440275
[22] J. Teng, X. Wang, J. Zhang, S. Wang, P. Huo, A multimodality medical image fusion
algorithm based on wavelet transform, In: Tan, Y., Shi, Y., Tan, K.C. (eds) Advances
in Swarm Intelligence. ICSI 2010. Lecture Notes in Computer Science, 6146. 2010,
pp. 627–633.
112 Data-Centric Artificial Intelligence for Multidisciplinary Applications

[23] B. Alfano, M. Ciampi, G. D. Pietro, A wavelet‑based algorithm for multimodal medical

image fusion, In: Falcidieno, B., Spagnuolo, M., Avrithis, Y., Kompatsiaris, I., Buitelaar,
P. (eds) Semantic Multimedia. SAMT 2007. Lecture Notes in Computer Science, vol
4816, pp. 117–120.
[24] X. Li, X. Tian, Y. Sun, Z. Tang, Medical image fusion by multi‑resolution analysis of
wavelets transform, In: Wavelet Analysis and Applications, Springer, 2007, pp. 389–396.
[25] Rajiv Singh, Ashish Khare, Multiscale medical image fusion in wavelet domain, The
Scientific World Journal, 2013, p. 521034, doi:10.1155/2013/521034
[26] Abhinav Krishn, Vikrant Bhateja, Himanshi, Akanksha Sahu, Medical image fusion
using combination of PCA and wavelet analysis, International Conference on Advances
in Computing, Communications and Informatics (ICACCI), Delhi, India, 2014,
pp. 986–991.
[27] H. B. Kekre, Archana Athawale, Dipali Sadavarti, Algorithm to generate Kekres wave‑
let transform from Kekres transform, International Journal of Engineering Science
and Technology, 2, 5, 2010, pp. 756–767.
[28] Bhakti Palkar, Dhirendra Mishra, Fusion of multi‑modal lumbar spine images using
Kekre’s hybrid wavelet transform, IET Image Processing, 13, 12, 2019, pp. 2271–2280.
[29] W. Selesnick, R. G. Baraniuk, N. G. Kingsbury, The dual‑tree complex wavelet trans‑
form, IEEE Signal Processing Magazine, 22, 6, 2005, pp. 123–151.
[30] J. Tao, S. Li, B. Yang, Multimodal image fusion algorithm using dual‑tree complex wave‑
let transform and particle swarm optimization, In: Huang, DS., McGinnity, M., Heutte,
L., Zhang, XP. (eds) Advanced Intelligent Computing Theories and Applications. ICIC
2010. Communications in Computer and Information Science, vol. 93. Springer, 2010.
pp. 296–303.
[31] Rajiv Singh, Richa Srivastava, Om Prakash, Ashish Khare, Multimodal Medical
Image Fusion in Dual Tree Complex Wavelet Transform Domain Using Maximum and
Average Fusion Rules, Journal of Medical Imaging and Health Informatics, 2, 2, 2012,
pp. 168–173.
[32] Pusit Borwonwatanadelok, Wirat Rattanapitak, Somkait Udomhunsakul, Multi‑focus
image fusion based on stationary wavelet transform, I. International Conference on
Electronic Computer Technology, Macau, China, 2009, pp. 77–81.
[33] R. Asokan, T. C. Kalaiselvi, M. Tamilarasi, Medical image fusion using stationary
wavelet transform with different wavelet families, Pakistan Journal of Biotechnology,
13, 2016, pp. 10–14.
[34] Wei Li, Xuefeng Zhu, Shaocheng Wu, A novel approach to fast medical image fusion
based on lifting wavelet transform, In: Proceedings of the 6th World Congress on
Intelligent Control and Automation, June 21–23, 2006, Dalian, China.
[35] Deron Ray Rodrigues, Curvelet based image fusion techniques for medical images,
International Journal for Research in Applied Science & Engineering Technology, 3,
III, 2015, pp. 310–316.
[36] M. N. Do, M. Vetterli, Contourlets, a directional multi‑resolution image representa‑
tion, Proceedings of IEEE International Conference on Image Processing, 1, 2002, pp.
I‑357–I‑360.
[37] Jyoti Agarwal, Sarabjeet Singh Bedi: Implementation of hybrid image fusion tech‑
nique for feature enhancement in medical diagnosis, Human‑centric Computing and
Information Sciences, 5, 2015, p. 3.
[38] T. Li, Y. Wang, Biological image fusion using a NSCT based variable‑weight method,
Information Fusion, 12, 2, 2011, pp. 85–92.
[39] H. Szu, I. Kopriva, P. Hoekstra, N. Diakides, M. Diakides, J. Buss, J. Lupo, Early
tumor detection by multiple infrared unsupervised neural nets fusion, In: Engineering
in Medicine and Biology Society, Proceedings of the 25th Annual International
Conference of the IEEE, vol. 2, IEEE, 2003, pp. 1133–1136.
Fusion of Multi-Modal Lumber Spine Scans 113

[40] Y.‑P. Wang, J.‑W. Dang, Q. Li, S. Li, Multimodal medical image fusion using fuzzy
radial basis function neural networks, In: International Conference on Wavelet Analysis
and Pattern Recognition, ICWAPR07, vol. 2, IEEE, 2007, pp. 778–782.
[41] W. Li, X.‑F. Zhu, A new algorithm of multi‑modality medical image fusion based on
pulse‑coupled neural networks, In: Wang, L., Chen, K., Ong, Y.S. (eds) Advances in
Natural Computation, Springer, 2005, pp. 995–1001.
[42] Sharma Dileepkumar Ramlal, Jainy Sachdeva, Chirag Kamal Ahuja, Niranjan
Khandelwal, Multimodal medical image fusion using non‑subsampled shearlet trans‑
form and pulse coupled neural network incorporated with morphological gradient,
Signal, Image and Video Processing, 12, 8, 2018. pp. 1479–1487.
[43] C.T. Kavitha a, C.Chellamuthu b, R. Rajesh, Multimodal Medical Image Fusion Using
Discrete Ripplet Transform and Intersecting Cortical Model, Procedia Engineering 38,
2012, pp. 1409–1414.
[44] Sudeb Das, M. K. Kindu, NSCT‑based multimodal medical image fusion using
pulse coupled neural network and modified spatial frequency, Medical & Biological
Engineering & Computing, 50, 10, 2012, pp. 1105–1114, doi:10.1007/s11517‑012‑0943‑3.
[45] Y. Liu, X. Chen, R. K. Ward et al., Image fusion with convolutional sparse representa‑
tion, IEEE Signal Processing Letters, 23, 12, 2016, pp. 1882–1886.
[46] Yu Liu, Xun Chen, Juan Cheng, Hu Peng, A medical image fusion method based on
convolutional neural networks, In: 20th International Conference on Information
Fusion Xian, China, 10–13 July 2017.
[47] https://www.image‑net.org/
[48] https://www.oreilly.com/library/view/advanced‑deep‑learning/9781789956177/b2258a
a6‑2c18‑449c‑ac00‑939e812f5a4a.xhtml.
https://www.oreilly.com/library/view/advanced‑deep‑learning/9781789956177/b2258a
a6‑2c18‑449c‑ac00‑939e812f5a4a.xhtml
[49] http://spineweb.digitalimaginggroup.ca
[50] Dhirendra Mishra, Bhakti Palkar. Article: Image fusion techniques: Review. International
Journal of Computer Applications, 130, 9, 2015, pp. 7–13​.
 8 Medical Image Analysis
and Classification
for Varicose Veins
Jyoti Yogesh Deshmukh, Vijay U. Rathod,
Yogesh Kisan Mali, and Rachna Sable

8.1 INTRODUCTION
Digital imaging and medicine are getting more and more integrated into society as
science and technology advance. Recent innovations in technology, such as image
processing and virtual reality, are rapidly finding their way into the medical industry.
The field of digital medicine has advanced quickly in recent years as interdisciplin‑
ary medical research combines with digital art.
The current healthcare system understands that while clinical procedures must be effi‑
cient, patient safety is of utmost importance. These conditions are not at all exclusive of
one another. For instance, when local or regional anaesthesia is used during surgery rather
than more intensive general anaesthesia, both patient safety and economic effectiveness
are maximized. It is still a difficult and open challenge to objectively evaluate the safety of
clinical procedures while taking patient comfort and financial expenses into account major
diseases in India caused by a bad lifestyle! Between 2005 and 2015, India’s proportion of
overweight and obese people increased. 21.7% of women and 19.6% of males between
the ages of 17 and 41 were not set in stone to be overweight or large. Corpulence is lethal!
Joint inflammation, malignant growth, barrenness, coronary illness, back agony, diabetes,
and stroke are the seven sicknesses most often connected to heftiness. Poor psychological
wellness, respiratory issues, irregular chemical characteristics, and food sensitivities are
also normal. Stoutness, which was once broad in major league salary nations, is currently
present and deteriorating in low‑ and centre‑pay nations. Obesity is mostly caused by an
inability to compensate for all of the energy consumed. Fat is created from this additional
energy. Fighting infectious diseases requires levying higher costs on unhealthy foods,
accurate labelling, and the development of environments that promote physical activity.
With a few notable exceptions, current therapies for obesity have largely demonstrated just
moderate success, and preventative strategies are relatively ineffectual. Therefore, a fuller
knowledge of the factors that contribute to severe or morbid obesity is urgently needed.
This understanding could result in novel and creative intervention strategies.
Accurate analysis of medical data will benefit in the early detection of illnesses,
patient care, and community services as machine learning in biomedicine and health‑
care quickly improves. The quantity and quality of unreported and inferior medical
data, on the other hand, jeopardizes disease analysis accuracy.

114 DOI: 10.1201/9781003461500-10

Medical Image Analysis and Classification for Varicose Veins 115

In order to identify obese people, this study provides an obesity detection algorithm
that is specifically designed for the Indian population. Additionally, based on the set of
symptoms entered into the system, it predicts which diseases the person is most likely
to get. Even if you feel healthy, it can tell you if you are at high or low risk. The system
certifies that a person’s health is within acceptable bounds and provides the option to
evaluate and monitor their health via email. Documents that patients (or maybe doc‑
tors) need to examine or recover as needed can also be uploaded and accessed easily.
The most prevalent peripheral vascular disease is varicose veins in the lower
limbs. There are now more than 25 million adults who have varicose veins in their
lower limbs, which affects about 23% of adults. Over 8% of Chinese people have
varicose veins. In addition to compromising aesthetics, varicose veins can induce
thrombophlebitis, venous oedema, cutaneous varicose veins, and an increased risk
of deep vein thrombosis. This could lead to disability and a decrease in employment
capacity. The direct annual cost of treating chronic venous disease in the USA is
estimated to be between $2.5 and $2 billion. Varicose vein treatment costs in the UK
make up 2% of all national health care spending. Conservative therapy and surgical
therapy are the two most common conventional treatments for varicose veins; how‑
ever, roughly one‑third of cases return within 10 years of surgery. Treatments with a
low level of invasiveness are being created. As a result, research into the molecular
and cellular mechanisms underlying fungal pathogenesis will be crucial in the future
for discovering novel treatment targets and creating novel therapeutic approaches.
Hemodynamic variations are directly sensed by vascular endothelial cells (VECs),
the unbroken monolayer layer that lines the lining of blood vessels between the blood
flow and the vessel wall. It is a mechanical barrier that serves a variety of important
roles in the body’s physiological and pathological processes, including substance trans‑
port, auto crime, and paracrime. In many illnesses, particularly in the lower extremities,
its structure and function are aberrant. It is essential for the growth and spread of vari‑
cose veins. The method proposed using harsh words in patients with varicose veins to
improve diagnosis accuracy through fundamental information processing [1]. Images
of vascular endothelial cells can be classified using this method based on various illness
situations. Ajitha [2] classified a number of noteworthy traits using a range of pattern
classifiers before using an artificial neural network to photos of leg veins. Shi et al.’s
[3] excellent classification of pictures of vascular endothelial cells using support vector
machines improved early detection of this disease. Zhu et al. [1] proposed employing
a grip sensation‑based responsive and predictive brain control method to investigate
lower extremity varicose veins. Veinidis et al. used 3D mesh sequences to construct an
unsupervised human motion search technique to assess whether lower extremities have
varicose veins. The aforementioned algorithm, meanwhile, has several drawbacks. B.
Vascular endothelial cell data are sparse, network training is challenging, and the adap‑
tive effect is weak. Through consecutive iterations of the training process, deep learn‑
ing uses a hierarchical feature extraction structure to abstract the input image from low
level to high level, identifying the most crucial file attributes. Extract more in‑depth
and broader features. In order to extract characteristics and classify them, this article
makes use of deep learning algorithms. In this study, lower extremity varicose veins
are detected and classified using features extracted from images of vascular endothe‑
lial cell inflammation using multi‑scale deep learning algorithms. Multiple convolu‑
tional layers were used to extract multi‑scale characteristics from photos of vascular
116 Data-Centric Artificial Intelligence for Multidisciplinary Applications

endothelial cells. In addition, we develop a competitive strategy that can lower the
network layer parameters while extracting more compact features by using the MFM
activation function rather than the ReLU activation function. For dimensionality reduc‑
tion, this network employs a method of 3 * 2 convolution kernel and 1 * 2 convolution
kernels. This can be applied to strengthen the network’s ability to extract features and
further optimize the network’s parameters.

8.1.1 Key Contributions

In this chapter, we have focussed on identifying varicose veins disease and its pos‑
sible data science methods to detect it. Some of the key contributions of this chapter
are as follows:

• How and why varicose veins affect the human legs?

• How it differs from normal veins of human legs?
• Use of multi‑scale technology to detect the stage or level of varicose vein
problem.
• Notable characterization and recognition of disease with deep convolutional
neural network (D‑CNN).
• Activation function and pooling methods are used to have fine categoriza‑
tion of the varicose veins.
• It has a detailed experimental set‑up with the use of the convolutional neural
network (CNN) for detecting the severity of the disease.
• Data science statistical methods are used to give evolutionary results of the
experiments.
• Metrics performance assessment will be done of all referred data science
statistical methods.

8.2 RELATED WORKS

To determine whether a lesion is present in the human body as well as to identify
and classify the lesion, doctors diagnose medical images. The programmed charac‑
terization and acknowledgement of clinical pictures have so collected a great deal
of consideration. Since the fiery peculiarity of vascular endothelial cells is firmly
connected with varicose veins in the lower furthest points, this article proposes a vas‑
cular endothelial cell‑based venous framework to accomplish the programmed order
and identification of provocative pictures and various pictures of varicose veins in
the lower limits. We offer a method for locating aneurysms. First, we photographed
the vascular endothelial cells in both healthy people and people with varicose veins.
Numerous convolutional layers are employed to extract multi‑scale information from
the image of vascular endothelial cells. Because of this, we extract more condensed
features using the max‑feature‑map (MFM) activation function as opposed to the
ReLU activation function, and we also incorporate a competing method to reduce
the network layer parameters. In order to further optimize the network parameters,
the network also uses a 1 * 2 convolution kernel for dimensionality reduction and a
3 * 2 convolution kernel to improve the network’s ability to extract features.
Medical Image Analysis and Classification for Varicose Veins 117

The clinical tissue classification system for identifying varicose veins using medi‑
cal imaging is suggested in this article. Over 90% of all cases of the many varieties
of varicose veins are ulcers. The system outlined in the proposed study can be used
to recognize and categorize varicose veins at any stage. The taxonomy is divided into
two sections in this article:
Texture features like homogeneity, energy, entropy, contrast, mean, dissimilar‑
ity, and variance are extracted during the feature extraction stage. To distinguish
between various twisting phases, these collected features are categorized using a
K‑nearest neighbour classifier and a support vector machine.
A significant fraction of people are affected by chronic venous insufficiency
(CVI), which cannot be treated without medical treatment. However, a lot of patients
don’t get their medical advice right away. Physicians also require methods to catego‑
rize patients based on the severity of their CVI at the same time. To help doctors and
patients, we suggest an automated categorization technique dubbed the CVI clas‑
sifier. An idea classifier in this strategy first guides low‑level picture highlights to
medium‑level semantic elements prior to building a multi‑scale semantic model to
develop a semantically rich picture portrayal. Second, a scene classifier is prepared
to gauge CVI seriousness utilizing the streamlined element subsets obtained by the
high‑request reliance‑based include determination approach. Classification perfor‑
mance is measured using the F1 score, kappa coefficient, and classification accuracy.
Routine outpatient varicose vein surgery is commonly performed under local
anaesthesia. Despite the fact that local anaesthesia is affordable and can reduce
patient risk, some patients endure discomfort during surgery. Thus, the careful group
should decide if to utilize general or neighbourhood sedation in light of an emotional
subjective evaluation of the patient’s nervousness and torment responsiveness. It is
absolutely impossible to confirm. To make sense of the relationship between cardio‑
vascular reaction change and agony during varicose vein a medical procedure, we
foster a three‑layered polynomial surface fitting of physiological information and
patient mathematical agony evaluations. Pulse changeability information was dis‑
sected for ghostly and underlying intricacy highlights as torment markers in no time
before 18 varicose vein systems. The aggravation expectation model that came about
was approved once, with a kappa coefficient of 0.82 (virtual arrangement) and a
region under the beneficiary working bend of 0.98 (close to consummate exactness).
This evidence of idea study shows the capacity of exactly evaluating torment aware‑
ness and science [4], allowing practitioners to prescribe the safest and least expensive
anaesthetic drugs to specific individuals in an unbiased manner.
By creating a cutting‑edge system to optimize machine learning algorithms for
accurately predicting obesity and related disorders among the Indian population,
this work seeks to address the aforementioned constraints. Unhealthy manufactured
foods are readily available, thanks to federal food restriction restrictions. More so
among younger generations, India’s increasingly career‑focused lifestyle has led to
an unpredictable biological pattern that favours indoor play over outside recreation.
The emergence of obesity nowadays is influenced by behavioural and socio‑psycho‑
logical factors as sleep, stress, race, and hormone imbalances [5].
Capillary blood pressure (CBP) is the main driver of fluid exchange between
microvessels. Prior to apparent peripheral oedema, asymptomatic systemic venous
118 Data-Centric Artificial Intelligence for Multidisciplinary Applications

congestion can have a direct impact on peripheral CBP. Therefore, in a range of

clinical diseases including heart failure, venous insufficiency, etc., CBP assessment
can provide early oedema control. At this time, CBP measurements can only be per‑
formed in invasive and difficult experimental circumstances [6]. We adjusted the fre‑
quently utilized oscillometric strategy in this study’s home blood vessel pulse screens
and fostered an opto‑mechanical gadget to empower harmless and programmed CBP
estimation. To quantify skin slender throbs, his proposed CBP framework is equipped
with a blue‑light photoplethysmogram sensor embedded in a finger/toe sleeve.

8.3 GAP ANALYSIS

Doctors diagnose medical images to evaluate whether a lesion is present in the
human body, as well as to identify and classify the lesion. There has been a great
deal of interest in the programmed grouping and acknowledgement of clinical pic‑
tures. Lower appendage varicose veins request programmed characterization and
distinguishing proof because of their immediate relationship to the fiery cycles of
vascular endothelial cells. This exploration made a historic hypothesis known as
MSD‑CNN, a varicose vein discovery strategy in view of pictures of vascular endo‑
thelial cell irritation and multi‑scale profound learning, to forestall such a phase of
clinical disappointment.

8.4 THE PROPOSED WORK

A. Overview of Lower Extremity of Varicose Veins Case
Varicose veins are now a highly frequent ailment as a result of society’s
ongoing advancement and improvements in living standards. The top layer
of skin is unaffected by human varicose veins in their early stages. Some
disorders are also readily disregarded by patients and receive little attention
from the general public. When venous valves in the legs are clogged for
an extended period of time, varicose veins can form Ref. [7]. Blood silting
from blood backflow leads to thrombosis, which happens as the condition
slowly spreads. After calf oedema and thrombosis, varicose veins in the legs
are brought on by a gradual darkening of the skin on the outer layer of the
calf. If medical attention is postponed, amputations and even cancers could
develop. A dangerous illness called thrombosis can significantly affect a
patient’s quality of life and length of life. Figure 8.1 compares varicose
veins and healthy veins.
Most vascular sicknesses are brought about by lower furthest point vari‑
cose veins, which are continuous venous illnesses in vascular medical pro‑
cedures. Eczema, a gradual loss of pigmentation, swelling, and distortion of
the superficial veins in the lower extremities – veins that are vulnerable to
chronic varicose veins, thrombophlebitis, and other consequences – are the
clinical signs. These complications can affect the patient’s ability to work
and lead a normal life has an adverse effect on.
Vascular endothelial cell inflammation is frequently a crucial connec‑
tion or expression of the onset of many cardiovascular, cerebrovascular, and
peripheral vascular disorders. One of the most prevalent vascular illnesses
Medical Image Analysis and Classification for Varicose Veins 119

FIGURE 8.1 Distinguish between varicose veins and normal veins.

is varicose veins of the lower extremities, which can cause leg oedema, dis‑
torted superficial vein dilation, severe varicose vein development, localized
bleeding, or infection. Varicose veins in the lower legs have been linked
favourably to vascular endothelial cell inflammation. The purpose of this
article is to identify and detect leg varicose veins by concentrating on vas‑
cular endothelial cells and extracting attributes from cell images.
B. Multi‑Scale Image Technology Method
Multi‑scale picture innovation is a strategy for communicating pictures
with different goals. The method involved with handling pictures at differ‑
ent scales is known as multi‑scale division after that.
Removing all highlights on a solitary scale in numerous visual imaging
applications could challenge. In this study, we present a multi‑scale tech‑
nique to facilitate feature extraction and improve feature extraction perfor‑
mance. The key tasks when employing multi‑scale picture technology are
expressing images in multi‑scale settings and figuring out how scales relate
to one another. A common example of a multi‑scale depiction of a picture
among them is an image pyramid.
Scale‑based analysis is a method that uses the picture pyramid. This
reflects the transform resolution for the identical image’s layers precisely.
With the biggest picture at the base and the littlest picture at the top, the
image size is commonly expanded and diminished in a pyramidal example.
Like the human visual framework, picture pyramid designs can address pic‑
tures at a few scales and depict the full picture at different scales. Pyramidal
pictures are developed utilizing down testing. Subsequently, the picture
order ascends as more picture detail is lost and the picture goal drops.
However, as demonstrated in Figure 8.2, photographs with lower quality
120 Data-Centric Artificial Intelligence for Multidisciplinary Applications

FIGURE 8.2 Image pyramid.

and size could include more generic elements. The resolution of vascular
endothelial cells, for instance, decreases with decreasing size and increases
in the sharpness of outlines and features following multi‑scale processing of
the pictures of vascular endothelial cells. Vascular endothelial cell images
get more detailed and greater in size as the resolution of the endothelial
cells increases. By removing features from multi‑scale images, image pyra‑
mid structures can achieve the goal of condensing multi‑scale information,
improving the relevance and depth of feature representation. Below is a
basic illustration of the many image pyramid types and structures.
1. Sub‑Sampling Pyramid
Image pyramids can be used to graphically portray multi‑scale struc‑
tures. Pyramiding is a common technique that involves applying a low
pass filter to smooth out an image and then down‑sampling the outcome.
One method for lowering the size of the image and improving the reso‑
lution of the two adjacent layers is down‑sampling. This method mostly
involves removing the horizontal and vertical pixels from the image.
Scale‑space theory states that the image size should be decreased using
the proper smoothing filtering. The resulting pyramid is a subsampled
pyramid if no filter smoothing is applied. An image that has been suc‑
cessfully downscaled using a subsampling pyramid has a much lower
resolution. To create thumbnail images that are only one‑fourth the size
of the actual image, the pixels and columns are primarily sampled pixel
by pixel for an image subsampling pyramid. This procedure is repeated
Medical Image Analysis and Classification for Varicose Veins 121

until the required number of photographs is obtained. The tiers of the

built‑up pyramid satisfy the experiment’s requirements.
The top and bottom images’ size ratios are always set to 1 when
creating an image pyramid. By down‑sampling the image of the subse‑
quent layer, the image of the preceding layer is produced. The following
organizing principles apply to the photo pyramid: The original image is
down‑sampled, and the layer with the lowest resolution is at the top, the
resolution of the layers behind it is down‑sampled, and the resolution
of the image continuously decreases from the bottom to the top. It has
the quality of being vulnerable. They are separated in half. The visual
organization is more pronounced in the photo pyramid as image resolu‑
tion decreases. With each lower level, the image’s resolution rises and
its local features and details are richly improved.
2. Gaussian Pyramid
Depending on the kind of filter employed, a low‑pass filter can be
used to smooth images when creating image pyramids. Pyramids of
many shapes and sizes are produced. An average filter, for instance,
results in an average pyramid. A Gaussian smoothing filter is used to
create a Gaussian pyramid. The quality of the subsampled image and
whether the preceding image’s pixels can accurately reflect the values
of the pixels in the subsequent level will determine which filter is used.
The basic principle behind the Gaussian Pyramid is to convolution the
base picture with a Gaussian kernel, test the subsequent picture to get
the past picture, and afterwards input the new picture to rehash the con‑
volution cycle with. The down‑inspecting approach is likewise done
various times to develop an information structure.
The expression for the Gaussian smoothing filter is as follows:

1 −( x 2 /2σ 2 )
g(x) = e (8.1)
2πσ
The standard deviation is the variable, and its magnitude affects how
smooth the signal is. Using the example of the 5 * 3 Gaussian kernels,
we may obtain the Gaussian kernel for convolution after discretizing the
Gaussian function.
The standard deviation formula is:

1
w= [1 2 6 4 16 24 6 24 36 4 1 16 4 24 6 ] (8.2)
256
3. The Pyramid of Gaussian Difference
The distinction of two Gaussian capabilities with contrasting fluc‑
tuations by the Laplace‑Gaussian, a band‑pass filter that reflects is an
approximation of the Gaussian difference. The filter function’s formula
is as follows:

D ( x , y, σ 1, σ 2 ) = ( G ( x , y, σ 1) ) − G ( x , y, σ 2 )) * I ( x , y ) = L ( x , y, σ 1) − L ( x , y, σ 2 )
(8.3)
122 Data-Centric Artificial Intelligence for Multidisciplinary Applications

C. The Deep Convolutional Neural Network (D‑CNN)

Artificial neural theory is the foundation of deep learning. It can focus
on nonlinear transformation of data, have a network structure with more
layers, and represent more abstract properties of images by optimizing the
learning method in comparison to typical neural networks. In the fields of
picture categorization and recognition, machine vision is widely used and
has achieved some notable recognition results. Speaking recognition, tar‑
get detection and tracking, computer advertising, recommender systems,
target classification, and image segmentation are just a few of the areas
where deep learning technology has achieved significant advancements.
The primary focal point of numerous universities, research foundations,
and organizations is innovation research. Profound learning can be uti‑
lized generally speaking where traditional man‑made consciousness inno‑
vations fizzle since it can advance significant level theoretical properties
from low‑level information, like picture pixels. Online directed and solo
learning procedures are utilized in profound learning. The profound con‑
volutional brain networks with start‑to‑finish learning capacities and solid
grouping and forecast execution are one of them, and they address admin‑
istered learning. An auto encoder is a portrayal of unaided learning and it
is equipped for remaking the first info picture. The highlights of the picture
are addressed by the centre organization layer. Unsupervised deep learning
networks select pictures that might not be used on the study day by giving
the network a lot of pre‑training with irrelevant pictures. Your informa‑
tion can be gotten to the next level. An immense number of picture assets
are presently accessible as a result of the Web and online entertainment,
which benefitted from the improvement of PC innovation and large infor‑
mation organizations. How much managed profound learning research has
developed? Individuals are especially worried about research on profound
convolutional brain organizations. Profound convolutional brain network’s
major areas of strength for models’ speculation execution, which has been
trained on databases with a size of over 1 million, accurately capture their
classification and discriminating abilities.
1. The neural network’s convolutional structure.
Convolutional brain organizations (CNN) are the most generally uti‑
lized profound learning network structure. Furthermore, it fundamen‑
tally affects various fields, including picture handling. When contrasted
with conventional picture handling techniques, CNNs offer the upside
of utilizing start‑to‑finish learning capacities to involve the real picture
as info directly and avoid image pre‑processing with little‑to‑no user
involvement. CNN offers the advantage of adopting local connection
architecture over standard neural networks, and it is distinguished by
simple model construction. The training weight parameter controls the
size of the convolution kernel because a weight distribution method is
employed to distribute weight across the neurons. Due to the fact that it
is unrelated to the quantity of neurons in the hidden layer, this further
limits the trainable parameters. The CNN network still has great feature
extraction abilities despite a significant decrease in the weight parameter.
Medical Image Analysis and Classification for Varicose Veins 123

Convolutional layers’ nodes are only hazily linked to one another

and to neighbouring network levels. A convolution kernel can extract
features. To add more features, one can use convolution kernels. A fea‑
ture map is the term used to describe this feature produced via convolu‑
tional kernel mapping. Usually, each layer is made up of several feature
maps. Each feature map is made up of several nodes that are squarely
organized. The same weight parameter or shared weight is inherited
by nodes with the same feature level. The initialization of a tiny value
matrix is typically randomized by convolution kernels during training.
The final training is given the right amount of weight thanks to ongo‑
ing learning. The likelihood that a CNN will over fit the training set
is decreased for each shared weight in a convolutional kernel that can
lower connectivity between neighbouring network layers. A down‑sam‑
pling layer is another name for a pooling layer. The two tactics of aver‑
age pooling and maximal pooling are widely used by him. The first one
is actually used. It has a wide range of applications. CNNs’ convolu‑
tional and pooling layers significantly lower the network’s complexity
and parameters. In a CNN, there are two components.
Convolution, activation, and pooling are methods used in feature
extraction and feature mapping. Finally, a softmax classifier is used to
categorize it. Figure 8.3 depicts the elements of a standard CNN struc‑
ture: a picture input layer, a convolutional layer, an enactment layer, a
pooling layer, a completely associated layer, a softmax classifier, and a
layer of convolutions.
Each understood unit can interface a restricted piece of the constant
district of the information picture in light of nearby identification limi‑
tations on associations among certain input units forced by the con‑
volutional layers of a convolutional brain organization. Convolutional
layer weight boundaries are comprised of a few convolution parts. In the
wake of playing out a joint convolution procedure on the whole picture,
a component map is extracted in a convolution check.

FIGURE 8.3 Structure of CNN.

124 Data-Centric Artificial Intelligence for Multidisciplinary Applications

Following is the formula to show how the convolution kernel

functions:
M

S ( i, j ) = ( X * W ) ( i, j ) + b = ∑( X
k =1
k + Wk ) ( i, j ) + b (8.4)

The quantity of information pictures is shown by the variable M,

where Xk is the kth picture and Wk is the kth convolution portion. The
variable b is the counterbalanced. A 3 * 4 convolution part and one
convolution step equivalent to one pixel are delivered when a 5 * 6 info
network is utilized. S(i, j) means the worth of the relating position com‑
ponent in the element map comparing to the convolution part W in the
component map S, where S is the element map relating to the convolu‑
tion portion W. Figure 8.4 shows the convolution process in action.
1. Activation Function
CNNs may tackle linear atomicity problems largely using
nonlinear formulas that are provided by activation functions.
CNNs include activation functions as a fundamental component.
Regardless of whether the organization order is a solitary layer
organization, the organization can direct straight planning with‑
out a trace of an initiation capability. Enactment capabilities can
be partitioned into two significant classifications: piecewise direct
capabilities and nonlinear capabilities with a dramatic construction
(Figure 8.5).
The most typical kind of activation function is the sigmoid func‑
tion with exponential shape, and its functional formula is:

1
f (x) = (8.5)
( e− x )
1 +

FIGURE 8.4 Schematic diagram of convolution operation.

Medical Image Analysis and Classification for Varicose Veins 125

FIGURE 8.5 Schematic diagram of sigmoid function.

FIGURE 8.6 Schematic diagram of ReLU activation function.

Another typical exponential shape activation function is the

Tanh capability (exaggerated digression), [−1, 1] is its scope and
capability articulation. The formula for the exponential shape acti‑
vation function is as follows:

( )(
f ( x ) = 1 − e −2 x / 1 + e −2 x ) (8.6)

The Tanh function exhibits soft saturation, a feature that is

similar to the sigmoid function and can easily lead to gradient dis‑
persion, the sigmoid function converges more slowly than it does,
though. The piecewise linear function of the ReLU linear correct‑
ing unit is written as:

f ( x ) = max ( x ,0 ) (8.7)

The chart of the ReLU enactment capability is displayed in

Figure 8.6. At the point when x > 0, ReLU can safeguard the slope
without lessening, settling the angle scattering issue and enabling
direct, supervised training of the deep network.
126 Data-Centric Artificial Intelligence for Multidisciplinary Applications

2. Pooling Layer
Sampling is the core of a pooling layer. In order to extract the
essential aspects of the input feature map while lowering the net‑
work parameters for improved characterization, the feature map
is somehow compressed. Decreased network parameters are also
helpful in avoiding over‑fitting.
There are two common pooling techniques. The best and typical
pooling. See the maximum 2 * 2 pooling with a 2‑pixel step size
in Figures 8.7 and 8.8. Only one of these pooling convolution part
loads can go over 1, while the rest can go under 0. The convolution
bit will be set where 1 signifies the most noteworthy worth of every
component on the element map.
In augmentations of two pixels, the convolution bit travels
through the component map. The component map is diminished to

FIGURE 8.7 Max pooling.

FIGURE 8.8 Mean pooling.

Medical Image Analysis and Classification for Varicose Veins 127

one‑fourth of its unique size subsequent to testing the best worth.

With the exemption that the normal of the things in the inclusion
zone of the component map is recovered and used as the result,
normal pooling is indistinguishable from the maximum example
procedure. The convolution piece’s sliding step size decides the size
of the last result map. The generated feature map will be one‑fourth
the size of the original image, using 2 pixels as an example.

8.5 RESULT AND DISCUSSION

A. Feature Extraction
Identifying the components that are most important for pattern identifi‑
cation is made possible by feature extraction, one of the most important pro‑
cedures in image processing. When a subset of characteristics enables the
creation of patterns with similarities within classes and differences between
other classes, it is considered to be a good subset of features. In the seg‑
mented image, varicose vein texture features are retrieved.
1. Texture characteristics
Texture refers to the spatial organization of colours or intensities
inside an image or specific regions of an image. We can presume the
texture of the image for natural situations that were caught in photo‑
graph picture segmentation is supported by the usage of picture textures
(Figure 8.9).

FIGURE 8.9 Block diagram of the experimental work.

128 Data-Centric Artificial Intelligence for Multidisciplinary Applications

• Homogeneity: The degree of dispersion can be measured using

the concept of homogeneity. It adjusts regional variances in image
textures while reflecting their local uniformity. A high uniformity
rating indicates that the local distribution of the image texture is
homogeneous, with no variance occurring within the area. The
Formula to calculate the degree of dispersion to measure homoge‑
neity is as follows:
N −1

∑ 1 + Pi(i −, j j )
i , j−0
2 (8.8)

• Energy: Energy can be thought of as an indicator of how often

pixel pairings are repeated. It checks the homogeneity of the image.
The energy value will be high if the pixels are highly similar. The
energy can be calculated using the following formula:
n−i

∑ ( Pi, j )
i , j−0
2
(8.9)

• Entropy: The texture of the input image is described by entropy, a

measure of randomness. When all components of the coexistence
matrix are equal, it’s worth is at its highest. The formula for entropy
calculation is as follows:
N −1

∑ − ln ( pi, j ) pij
i , j=0
(8.10)

• Contrast: Finding the intensity difference between a pixel and

its neighbours throughout the image is known as contrast, used to
enhance the quality of images. The variation in hue that makes an
object stand out is called contrast. Differences in hue and bright‑
ness between things and other objects in the same field of vision
are visual indicators of contrast. The formula to find the contrast
between pixels is as follows:
N N

∑∑ (i − j ) ( P (i, j ))
i j
2
(8.11)

• Mean: By averaging every pixel value in the image, average is

determined. The average values show how each pixel’s intensity
contributed to the overall image. The mean of every pixel will be
calculated with the following formula:
N N

µ= ∑ j ( p) = ∑i ( p)
j =1
ij
i =1
ij
(8.12)
 Medical Image Analysis and Classification for Varicose Veins 129

– Dissimilarity: The numerical scale for comparing the similarity

or difference of two data from 0 (objects are alike) to ∞ (objects
are different). To calculate dissimilarity, the following formula is
referred:
N N

∑∑Pi =1 j =1
ij i− j (8.13)

– Variance: Variance is typically used to categorize images into dis‑

tinct regions by determining how each pixel differs from the centre
or nearby pixels. The categorization of images will calculate with
variance and its formula is as follows:
N N

V = Vi ∑∑ (1 − µi )
i =1 j =1
2
Pi , j (8.14)

B. Metrics for Performance Assessment

The proportion of accurately diagnosed varicose vein stages, or True
Positives, is used to gauge the suggested method’s sensitivity. It is calculated
with the following formula:
Sensitivity No of cases matched with clinicians report
= ∗ 100
No of cases matched with clinicians report + No of cases not matched with clinicians report
(8.15)

Utilizing predictions of false positives, specificity is evaluated. The per‑

centage of varicose vein stages that the technology detected but the doctor
did not report is known as the false positive rate. Rarely may clinicians mis‑
interpret the correct stage with the aid of specificity. Specificity is evaluated
with following formula:

No. of cases identified by the system

Specificity = 1 − *100 (8.16)
No. of cases reported by the clinicians

Accuracy is a measure of a system’s ability to anticipate outcomes in

terms of genuine positive, genuine negative, misleading positive, and mis‑
leading negative values. The formula to calculate system’s ability accuracy
is as follows:

True Positive+True Negative

Accuracy = (8.17)
True Positive+True Negative+False Positive+False Negative

The suggested system is measured using these three separate metrics.

Thirty‑three photos in total were needed to train the system. Thirty differ‑
ent photographs were utilized in a test set that was created using a random
sampling technique (Tables 8.1 and 8.2).
130 Data-Centric Artificial Intelligence for Multidisciplinary Applications

TABLE 8.1
Confusion Matrix for the CNN Classifier’s Test Set
Test Set Genuine Positive Misleading Positive Genuine Negative Misleading Negative
T1 14 7 34 6
T2 15 6 33 5
T3 14 7 35 7

TABLE 8.2
Statistical Performance Analysis of CNN and SVM Classifier
CNN Classifier SVM Classifier
Test set Sensitivity Specificity Sensitivity Specificity
T1 0.772929 0.764768 0.896825 0.884729
T2 0.888889 0.955555 1 0.926833
T3 0.823576 0.968254 0.552632 1

FIGURE 8.10 The pictorial sensitivity specificity analysis of KNN and SVM classifier.

The results of the testing phase were used to calculate the aforemen‑
tioned performance metrics. Figure 8.10 shows the image sensitivity spec‑
ificity analyses for the CNN and SVM classifiers. Table 8.3 presents the
findings of the accuracy analysis. Experiment results reveal that the SVM
classifier offers sensitivity (77.94%), specificity (86.78%), and accuracy
(78.84%) while the CNN classifier of the suggested approach offers sensi‑
tivity (72.46%), specificity (85.17%), and accuracy (80.95%).
Medical Image Analysis and Classification for Varicose Veins 131

TABLE 8.3
Accuracy Analysis of KNN and SVM Classifier
CNN Classifier SVM Classifier
Test set Accuracy Accuracy
T1 0.869542 0.845654
T2 0.936488 0.92238
T3 0.918635 0.801268

8.6 CONCLUSION
Now more than ever, early detection and treatment of varicose veins require com‑
puter‑assisted image analysis for histological classification. The suggested model
extracts textural information that is extremely helpful for classifying different
phases of wounding. He also uses two classifiers in this strategy; thus, the efficacy of
both classifiers will aid in the discovery of a more accurate classifier in subsequent
research. The results for the classifier to identify varicose vein stages will be increas‑
ingly accurate as more datasets and processing steps are added to the approach in
the future.

REFERENCES
1. Zhu, Ruizong, Huiping Niu, Ningning Yin, Tianjiao Wu, and Yapei Zhao. “Analysis
of varicose veins of lower extremities based on vascular endothelial cell inflammation
images and multi‑scale deep learning.” IEEE Access 7 (2019): 174345–174358.
2. Ajitha, K. “SVM VS KNN for classification of histopathological images of varicose
ulcer”, Advances in Engineering: an International Journal (ADEIJ) 3, no. 2 (2020):
19–28.
3. Shi, Qiang, Weiya Chen, Ye Pan, Shan Yin, Yan Fu, Jiacai Mei, and Zhidong Xue.
“An automatic classification method on chronic venous insufficiency images.” Scientific
Reports 8, no. 1 (2018): 17952.
4. Barulina, Marina, Askhat Sanbaev, Sergey Okunkov, Ivan Ulitin, and Ivan
Okoneshnikov. “Deep learning approaches to automatic chronic venous disease clas‑
sification.” Mathematics 10, no. 19 (2022): 3571.
5. Adjei, Tricia, Wilhelm Von Rosenberg, Valentin Goverdovsky, Katarzyna Powezka,
Usman Jaffer, and Danilo P. Mandic. “Pain prediction from ECG in vascular surgery.”
IEEE Journal of Translational Engineering in Health and Medicine 5 (2017): 1–10.
6. Pereira, Naomi Christianne, Jessica D’souza, Parth Rana, and Supriya Solaskar.
“Obesity related disease prediction from healthcare communities using machine learn‑
ing.” In: 2019 10th International Conference on Computing, Communication and
Networking Technologies (ICCCNT), pp. 1–7. IEEE, 2019.
7. Liu, Jing, Bryan Yan, Shih‑Chi Chen, Yuan‑Ting Zhang, Charles Sodini, and Ni Zhao.
“Non‑invasive capillary blood pressure measurement enabling early detection and clas‑
sification of venous congestion.” IEEE Journal of Biomedical and Health Informatics
25, no. 8 (2021): 2877–2886.
 9 Brain Tumor Detection
Using CNN
Paras Bhat, Sarthak Turki, Vedyant Bhat,
Gitanjali R. Shinde, Parikshit N. Mahalle,
Nilesh P. Sable, Riddhi Mirajkar,
and Pranali Kshirsagar

9.1 INTRODUCTION
The human body is a combination of interrelated parts or networks where each part
is interconnected to the other, and dysfunction in one part shows the impact on the
overall body of the individual. Being such a complex system nature has provided it
with an inbuilt processor which manages all its work and responds to every stimulus
in a reasonable manner. Humans have named it as brain, the most important organ
of human beings. Our existence is immensely dependent on the proper functioning
of the brain that performs most of our tasks, be it controlling voluntary movement,
creating and managing memories, developing thoughts, etc. [1].
Now‑a‑days, due to processed foods, the use of plastics in our day‑to‑day life, the
consumption of adulterant drinks, the increase in smoking among youth, etc., have
made this disease of cancer spread like a forest fire which is increasing at a very fast
pace [2]. Some of the brain cancers are depicted in Figure 9.1.
A brain tumor is such a kind of tumor in which the tissues inside the brain start to
grow abnormally creating an extra piece of mass inside the brain which takes away
the nutrients of its surrounding cells, thus resulting in brain failure. The disease is
curable if it is identified at an early stage, which is the most challenging part of this
process as it mildly shows any symptoms at its early stage and often gets skipped
away by doctors. In order to help the live saviors, i.e. doctors to predict this problem
at an early stage, here this project aims to provide some more time for the doctors to
think and curate this problem as early as possible. We have tried to build a machine
that uses CNN and many more algorithms to detect the brain tumor at its early stage.
Our machine identifies a tumor from a picture and returns the result of whether the
tumor is positive or negative, which makes it useful in situations when we need to be
certain of the tumor status.
The primary goal of brain tumor detection is to classify various tumor forms in
addition to just detecting them. It also serves the purpose of identifying a tumor from
a picture and gives the output of whether the disease is present or not, which makes it
useful in situations when an infected person needs to be certain of the tumor’s status.
This project is focused on developing a system that can recognize tumor blocks in
MRI scans of various patients and prevent the damage it causes in a patient’s life.

132 DOI: 10.1201/9781003461500-11

Brain Tumor Detection Using CNN 133

FIGURE 9.1 Types of brain cancer [3].

9.2 LITERATURE SURVEY

Sasikala et al. [4] suggested some new self‑activating tumor detection methods with the
help of Deep Neural Network (DNN) for proficient glioblastoma detection. It uses a final
layer that implements fast segmentation on the order of 24 s–3 min across the lung area.
Jyothi et al. [5] proposed eight different kinds of MRI scans out of which seven represent
different tumor kinds and one represents basic brain tumor. The method of Deep CNN
and SVM uses it in a collection of standard data items intended to facilitate systematic
measurement comparison in order to achieve a precision of 92.17% with a correctness of
93%. Joshi et al. [6] focused on brain tumor segmentation. Only quantitative measures of
disease modeling allow process monitoring and recovery. The model is more susceptible
for detecting defects related to brain and stroke, brain tumors or infections.
Kiranmayee et al. [7] proposed a prototype which consists of both testing and
training phases and implements the enhancement of service by combining health and
network support emotionally in the area of health service and improving its quality.
Arya et al. [8] have given an overview of different image preprocessing and division
methods, which contain picture filtering methods, noise‑removing methods, usage of
graphs and algorithms like watershed. Siar et al. [9] proposed the method of CNN
134 Data-Centric Artificial Intelligence for Multidisciplinary Applications

while reaching an accuracy of 98.66% and also used the radial basis function and
Decision Tree classifier. The SoftMax Fully Connected Plate used for image clas‑
sification had a classification accuracy of 98.57%.
Deepak et al. [10] proposed the methods that were used: GoogleNet and CNN, and he
also describes the advantage of CNN‑based classifier systems is that they do not require
manually segmented tumor regions and provide a fully automated classifier. Demiharan
et al. [11] suggested a segmentation technique for categorizing brain tumor MRIs. Using
station wavelet transform, learning vector quantization, cerebral spinal fluid (CSF),
edema, white matter (WM), and grey matter were on the order of 0.87 for grey matter,
0.96 for CSf, and 0.77 for edema. WM was discovered in 0.91%. Aneja et al. [12] sug‑
gested a segmentation algorithm that uses fuzzy C‑means (FCM) clusters for noise fig‑
ures as well as a fuzzy clustering averaging technique. The cluster validity function, run
time, and convergence error rate of 0.537% are used to evaluate segmentation values.
Yang et al. [13] used discrete wavelength transform (DWT) with an accuracy of
93.9% and an objective prediction of 6.9%. Badza et al. [14] proposed their own
CNN architecture for three types of brain tumor classification. The proposed model
is more straightforward than the existing pre‑trained models. They used T1W‑MRI
data for the training and testing with tenfold cross‑validation.
These approaches have suggested many ways in which the model has achieved efficient
ways to diagnose a brain tumor. Table 9.1 summarizes the literature work on brain cancer.

9.3 PROPOSED ALGORITHM

The brain being the most important and delicate part is also the most complex organ of
the body of a human. Understanding the functioning of the brain in itself is a tedious task,
but through our project, we have tried to understand its complex behavior and developed
a project which uses convolutional neural network architecture to detect the region accu‑
mulated by tumor by processing its MRI images, that too at an early stage. The main pur‑
pose is to help doctors in predicting diseases at an early stage and save the lives of people.
The working of the CNN model is shown in Figure 9.2; details of the CNN layers
are as follows:

• The Convolution Layer: The main layer of CNN is a convolution layer

which consists of kernels or filters present in it whose size is smaller than the
actual image. The convolution layer has the function of extracting the fea‑
tures of the input image given to it and returning the output in matrix form.
• Activation: The activation layer consists of an activation function inside the
layer. For our model, we have chosen the activation function ReLU. The pur‑
pose of ReLU activation function is that it will give the output as negative
if the input is negative and will give the same output if the input is positive.
• Max Pooling: The max pooling layer reduces the dimensions of the image by
taking out the largest element present in the matrix on which pooling is used.
• Flatten: Flatten is another layer present in CNN which is used to convert
the pooled feature matrix into a list. The output given by the flatten layer is
taken as the input to the fully connected neural network.
• Dense: The dense has the function of connecting the fully connected layer
to the neural network.
 Brain Tumor Detection Using CNN
TABLE 9.1
Summary of Literature Work on Brain Cancer
Research
work Used Methods Used Dataset Obtained accuracy Advantages Discussion
Wentao Methods used were SVM BraTS 2014 and CNN 87.57% and SVM 86.04% Faster segmentation Need improvement in
et. al. [1] and deep CNN BraTS 2016 accuracy
S et al. [4] Methods used are neural BraTS 2010 FCM in WM, GM, CSF 30.01, 31.04, Enhancement of picture with noise Need of updation in
networks, K‑means and 28.04 that too with least error misclassification error
fuzzy logic
Jyoti Methods used are deep OASIS 92.17% Precision, 93% correctness, Multi‑class classification shows Size of the model is
et al. [5] CNN and SVM 92% recall and 91% f1‑score great significance evanescently small and
can’t handle big datasets
Joshi Methods used are image 30 research HSOM scans 110 abnormal and 62 Problem of image restoration and Algorithm could solve
et al. [6] segmentation, restoration papers from normal axial MRI images with a enhancement of images has been particular research
and enhancement of 2000–2015 92.41% accuracy resolved and explained with proper problems only.
images methodologies.
Kiranmayee Detecting brain tumor The outcomes gained demonstrate It was on a prototype
et al. [7] consisting of training and that the combination of stage
testing phase emotionally supportive networks
with medical services can improve
the quality of services.
(Continued)

135
 136
Data-Centric Artificial Intelligence for Multidisciplinary Applications
TABLE 9.1 (Continued)
Summary of Literature Work on Brain Cancer
Research
work Used Methods Used Dataset Obtained accuracy Advantages Discussion
Deepak Methods used were Figshare DCNN 91.2% and SVM Classifier Stable and efficient Lack of accuracy in
et al. [10] GoogleNet and deep CNN 0.98% transfer model
Demirahan Methods used are neural IBSR 2015 and WM 90%, GM 87%, hydrops 76%, Enhancement of efficiency in Can’t be applied on
et al. [11] networks, self‑organizing BraTS 2012 tumor 60% and CSF 95% WM.GM and edema newly generated dataset
maps and Wavelets
Aneja Method used is fuzzy NSL‑KDD FCM 1.173, T2FCM 0.951 and IFCM Reduction of disturbances in Need of updation in
et al. [12] clustering mean algorithm 0.436 training set & size clot misclassification error
Yang Methods used is DWT GE Healthcare Collected reliability of 93.9% and an More work on deduction on SVM Crises of Model handling
et al. [13] objective error rate of 6.9%
Badza Method used is CNN BRATS and Repeating the Fitting procedure 10 Could be used for differential Heavy run‑time
et al. [14] CBICA times we get the punctuality of 95.08% datasets
Brain Tumor Detection Using CNN 137

FIGURE 9.2 CNN model working for brain tumor detection.

FIGURE 9.3 Architecture of CNN model.

The dataset is taken from github, 253 MRI images with 155 positive instances and 98 neg‑
ative samples. The neural network couldn’t be trained because the dataset was too little. In
order to address the problem of data imbalance, data augmentation proved helpful. Data
augmentation is used to increase the dataset. The dataset now has 1085 positive examples
and 980 negative examples, for a total 2065 example photos, following data augmentation.
The following pre‑processing processes were used for each image:
Images are resized such that they only show the brain in one section (which is the
most important part of the image). The neural network accepts images of the same
sizes, hence image shapes are kept as 240, 240, 3 = (image width, image height, number
of channels). The value of pixels can be scaled using Normalization to a range of 0–1.
The data were divided as follows:

• Training was done with 70% of the data

• To validate, 15% of the data
• Test will use 15% of the data

9.3.1 Neural Network Architecture

The system architecture is shown in Figure 9.3; let’s discuss each layer of the archi‑
tecture in detail:
138 Data-Centric Artificial Intelligence for Multidisciplinary Applications

FIGURE 9.4 MRI images without tumor.

The first step involves giving the input to the neural network. The neural network
is given an input image with a shape of (240, 240, 3) for each input image x. When
the image is given as input to the neural network, it traverses the following layers:
The first layer it traverses is the zero‑padding layer which is of the size of (2,2).
Then followed by the zero‑padding layer, there is a convolutional layer which
consists of 32 filters, with a stride of 1 and the filter with the size of (7,7). After
the convolution layer, there is a normalization layer which helps in normaliz‑
ing the pixel values in a batch to speed up computation. After the normalization
layer, there is the activation layer which consists of the activation function. The
­activation function we used is the ReLU activation function. After the activa‑
tion layer, there is a max pooling layer with a filter size of 4 and a stride of 4.
Then again, there is an identical layer of Max Pooling with f = 4 and s = 4. After
the pooling layers, there is a flatten layer which converts the three‑dimensional
matrix into a vector with only one dimension. After the flatten layer, there is a
dense layer in which one neuron is in a dense, fully linked layer with an output
unit that has sigmoid activity.
MRI images without brain tumor and having brain tumor are shown in Figures 9.4
and 9.5, respectively.

9.4 RESULT AND DISCUSSION

The result was generated through continuous testing of the model which was
done as follows: First, load the model and we are doing that by using the file path:
model = load_model (filepath=‘models/cnn‑parameters‑improvement‑23‑0.91.model’).
So we have a model.metrics_names, the purpose at the time of compilation is to
check against the monitored quantities which are very important in the callback. To
see if the model is performing well, it becomes essential to evaluate it against the
test model. A test model is based upon a pre‑processed dataset using data prepara‑
tions. We obtain the correctness based on the F‑measure on the test model, and the
result we achieved is: For loss, it is 11.3%, and for the accuracy, it is 88.7%, f1‑score
is 0.88.
Brain Tumor Detection Using CNN 139

FIGURE 9.5 MRI images with tumor.

Let’s keep in mind the ratio of positive to negative examples: Start by declaring
the variables m and n_positive for the size of the dataset and the quantity of positive
examples, respectively. Now, we can compute the total examples that are negative
i.e. n_negetive=m‑n_positive

Number of examples: 2065

Positive Examples as Percentage: 52.54237288135593%
Number of positive examples: 1085
Negative Examples as Percentage: 47.4576271186440%
Number of negative examples: 980

Training Data:

So, the Percentage of Positive and Negative data is 52.8719723183391%

and 47.1280276816609% respectively while the Number of Positive and
Negative samples are 764 and 681

Validation Data:

Number of Examples: 310

Consequently, the ratio of positive to negative samples is from
54.83870967741935% to 45.16129032258065%.
We get 170 positive samples and 140 negative samples

Testing Data:

Number of Examples: 310

Positive samples as a Percentage is 48.70967741935484% while as a number it
is 151. Negative Sample as a percentage is 51.29032258064516% while as a
number it is 159
140 Data-Centric Artificial Intelligence for Multidisciplinary Applications

FIGURE 9.6 Graph of training and validation accuracy.

FIGURE 9.7 Graph of training and validation loss.

Graphs shown in Figures 9.6 and 9.7 show that with the increase in the testing data
of the model, its training and validation accuracy is increasing continuously while
the training and validation loss is being continuously reduced which is shown in
Figures 9.6 and 9.7.
Brain Tumor Detection Using CNN 141

9.5 CONCLUSION
In this work, we reviewed the available feature‑based research in the literature. We
have implemented the CNN model with 88.7% accuracy on the test set and a score
of 0.88 on the test set for f1. This model can detect brain cancer, and the results are
satisfactory when you consider how balanced the data is. This will help the doctors
to predetermine the disease and save more lives.

REFERENCES
1. Wu, Wentao et al. (2020). An intelligent diagnosis method of brain MRI tumor segmen‑
tation using deep convolutional neural network and SVM algorithm. Computational
and Mathematical Methods in Medicine, Volume 2020, Special issue.
2. Deepa, S. A. (2016) Review of brain tumor detection from tomography. In: International
Conference on Computing for Sustainable Global Development (INDIACom).
(pp. 3997–4000). IEEE.
3. Siar, M., & Teshnehlab, M. (2019, October). Brain tumor detection using deep neural
network and machine learning algorithm. In 2019 9th international conference on com‑
puter and knowledge engineering (ICCKE) (pp. 363–368). IEEE. https://www.zeeva.in/
know‑everything‑about‑brain‑cancer/
4. Sasikala, S., Bharathi, M., & Sowmiya, B. R. (2018). Lung cancer detection and classifi‑
cation using deep CNN. International Journal of Innovative Technology and Exploring
Engineering, 8(25), 259–262.
5. Jyothi, P., & Singh, A. R. (2022). Deep learning models and traditional automated tech‑
niques for brain tumor segmentation in MRI: a review. Artificial Intelligence Review,
56(4), 2923–2969.
6. Joshi, D., & Goyal, R. (2017). Review of tumor detection in brain MRI images. In:
2019 International Conference on Innovative Trends and Advances in Engineering and
Technology (ICITAET) (pp. 206–209). IEEE.
7. Kiranmayee, B. V., Rajinikanth, T. V., & Nagini, S. (2017, September). Explorative
data analytics of brain tumour data using R. In: 2017 International Conference on
Current Trends in Computer, Electrical, Electronics and Communication (CTCEEC)
(pp. 1182–1187). IEEE.
8. Arya, M., & Sharma, R. (2016). Brain tumor detection through MR images: a review of
segmentation techniques. International Journal of Computer Applications, 975, 8887.
9. Siar, M., & Teshnehlab, M. (2019). Brain tumor detection using deep neural network
and machine learning algorithm. In: 2019 9th International Conference on Computer
and Knowledge Engineering (ICCKE) (pp. 363–368). IEEE.
10. Deepak, S., & Ameer, P. M. (2019). Brain tumor classification using deep CNN features
via transfer learning. Computers in Biology and Medicine, 111, 103345.
11. Demirhan, A., Törü, M., & Güler, I. (2014). Segmentation of tumor and edema along
with healthy tissues of brain using wavelets and neural networks. IEEE Journal of
Biomedical and Health Informatics, 19(4), 1451–1458.
12. Aneja, D., & Rawat, T. K. (2013). Fuzzy clustering algorithms for effective medical image
segmentation. International Journal of Intelligent Systems and Applications, 5(11), 55–61.
13. Yang, G., Nawaz, T., Barrick, T. R., Howe, F. A., & Slabaugh, G. (2015). Discrete wave‑
let transform‑based whole‑spectral and subspectral analysis for improved brain tumor
clustering using single voxel MR spectroscopy. IEEE Transactions on Biomedical
Engineering, 62(12), 2860–2866.
14. Badža, M. M., & Barjaktarović, M. Č. (2020). Classification of brain tumors from MRI
images using a convolutional neural network. Applied Sciences, 10(6), 1999.
 10 Explainable Artificial
Intelligence in the
Healthcare
An Era of Commercialization
for AI Solutions
Prasad Raghunath Mutkule, Nilesh Popat Sable,
Parikshit N. Mahalle, and Gitanjali R. Shinde

10.1 INTRODUCTION
The significance of explainable artificial intelligence (XAI) is a term encompassing
techniques that render AI systems interpretable and comprehensible to end‑users. The
importance of XAI is underscored by its applicability in numerous fields. In areas
such as healthcare, ensuring the reliability of AI models is paramount. Furthermore,
AI explainability can lead to new insights in fields such as Physics, Mathematics,
and Chemistry. Equally important is the need for everyday users of AI to understand
its decision‑making processes. Additionally, XAI can aid neuroscience research
in the testing and explanation of hypotheses related to brain activity and learning
mechanisms. Apart from explicating the importance of XAI, this chapter delves into
two XAI methods – model‑agnostic and model‑specific. The universal applicability
and sensitivity analysis feature of model‑agnostic methods positions them as suit‑
able for any ML model. Conversely, model‑specific methods are tied to certain ML
models, employing techniques such as activation maximization and deep network
understanding and visualization. XAI solutions adhere to specific criteria, such as
trustworthiness, transferability, causality, and interactivity, that provide a framework
for assessing the quality and efficiency of XAI methods. Through this investigation
of XAI and its significance, it seeks to demystify AI systems, making them more
transparent and interpretable.
Machine learning (ML) algorithms can produce findings that are understandable
to humans with the help of XAI. An XAI model predicts an impact and its biases.
It assists in decision‑making aided by AI by evaluating correctness, fairness, and
transparency. A company’s ability to explain AI is crucial when it comes to bring‑
ing AI models into production. A responsible AI development strategy can also be

142 DOI: 10.1201/9781003461500-12

Explainable Artificial Intelligence in the Healthcare 143
XAI overview.
FIGURE 10.1
144 Data-Centric Artificial Intelligence for Multidisciplinary Applications

implemented when AI is explainable. The ability to understand and retrace how a

computer arrives at a conclusion will become more difficult as AI advances. A “black
box” is created to represent the entire mathematical process, which makes it difficult
to understand. These black‑box models are constructed using data [1]. A lot of times,
even the people who created the algorithm don’t understand what’s going on inside
them or how it came up with a particular conclusion [2]. According to the business
world, XAI identifies patterns and predicts desirable behaviours that lead to improved
outcomes. With XAI, the firm owner can directly manage AI operations, since he
already understands what the machine does. By understanding and understanding
how AI systems work, stakeholders can build trusting relationships with you. It also
ensures the company’s safety, as all operations should follow safety regulations. It
takes commitment to comply with new security regulations. Under the existing Right
to Justify statute, all choices taken quickly are prohibited. Figure 10.1 gives an over‑
view of XAI.

10.2 PRINCIPLES OF EXPLAINABLE AI

AI systems must consider how the information will be received by humans as they
formulate these principles [3]. All factors will affect which explanation is appropri‑
ate for a given situation, including the requirements of the situation, the task at hand,
and the consumer. Regulations and laws, quality control of AI systems, and customer
relations are a few examples of these situations. Motives, reasons, and perspectives
are captured by the four principles.
It is important to note that XAI is based on four principles:

10.2.1 Explanation
There is an explanation or reasoning accompanying all outputs. AI systems must
provide proof, support, and rationale in support of each output under the Explanation
principle. Essentially, this theory implies that an explanation can be provided by a
system without it necessarily being justified, appropriate, or instructive in and of
itself; the only requirement is that the system can provide a rational explanation [4,5].
In the current state of technology, this type of XAI technique is being developed
and validated extensively. There are a number of technologies and methods being
developed and implemented at the moment. Their explanations are not subject to any
quality criterion.

10.2.2 Meaningful
System explanations are understandable by individual users. A meaningful system is
one that is understood by the recipient of its explanations. A user meets this concept
when the explanation is understood by them and/or helpful to them in completing a
task. A one‑size‑fits‑all approach may not be the best answer based on this principle.
It may be necessary to provide different explanations to different groups of users for
the same system. Users may receive explanations tailored to their needs due to the
Meaningful principle [6–8]. There are many large groups of people in the world,
Explainable Artificial Intelligence in the Healthcare 145

such as developers versus users, lawyers versus judges, etc. There may be some dif‑
ferences between the objectives and desires of these two groups. The importance of
certain factors may differ between forensic practitioners and juries. It is also possible
to personalize explanations for individuals using this concept as well. It is found that
there is sometimes a difference in perception between individuals who are observing
the output of the same AI system, for a variety of reasons.

10.2.3 Explanation Accuracy
It is accurate to describe the output generation process of the system in the descrip‑
tion. Explanations that are meaningful to users can only be generated by systems
if they are applied together with the Meaningful principle. System output genera‑
tion processes are not required to be accurately explained in order to comply with
these two principles. Providing accurate explanations is essential to the Explanation
Accuracy principle. Correctness of explanation is different from decision accuracy.
Decision accuracy refers to the system’s ability to make the right decision when mak‑
ing decisions. Although the system may make an accurate judgement regarding the
situation, the accompanying explanation may not accurately explain how the results
were reached, no matter how accurate the judgment may be [9]. It has been estab‑
lished by AI researchers to create standard metrics that measure the accuracy of
algorithms and systems. It is understood that there is no performance metric that can
be used to measure the accuracy of explanations, although there are reliable measure‑
ment methods available.

10.2.4 Knowledge Limits

In order for a system to work properly, it has to operate under conditions that lead
to the output of the system being reliable under those conditions. System boundaries
are implied by the preceding concepts. In light of this Knowledge Limits concept, it
is argued that systems are capable of detecting situations that are beyond the scope
of their design or approvals. In this approach, we identify and express limitations
of knowledge so that we don’t make judgements when we don’t need to. With the
Knowledge Limits Principle, you can build trust in a system by eliminating decep‑
tive, harmful, or unfair decisions or outcomes. One of two approaches can be taken
when dealing with a system’s knowledge limitations [10–12]. First, it could be that the
inquiry isn’t within the purview of the system. Users can upload pictures of apples to
bird species categorization systems, for instance. If the system was unable to locate
any birds in the supplied image, then it might provide an answer indicating that it
was unable to provide an answer because it could not locate any birds in the image.
I am going to respond to your question as well as explain my reasoning [13]. Based
on a threshold of internal confidence, one may be able to determine which of the
two methods has the most likely response, but the second method may have reached
a knowledge limit in working with a limited amount of data. For example, a bird
classification system may not be able to establish the species of a bird based solely
on the input image of the bird. Despite the fact that this image is of poor quality, the
algorithm may still be able to recognize that it is of a bird in this example [14,15].
146 Data-Centric Artificial Intelligence for Multidisciplinary Applications

10.3 NEED OF AI FOR EXPLANATION

There are some AI systems that do not require explanation. Occasionally, we will
need an explanation only in cases where we are the ones who require it. Making a
system that requires an explanation requires knowledge of which system needs one.
Developers can then use a strategy that is appropriate for their situation [16]. In order
to figure out when our system must explain something, we need to take into account
the following points listed below:

10.3.1 The Need for Fairness

There are certain forms of fairness within the system that are mandatory, and people
cannot compromise when it comes to fairness. Statistical predictions are critical to
the success of any project and are able to have an impact that is irreplaceable and
lasting in the long term. Depending on the situation, a physician may recommend
an operation, suggest receiving hospice care for the patient, etc. The system has the
obligation to explain when there is a high cost associated with a mistake, where
the system is required to produce the right result in response to the mistake [6,17]. In
the event it is incorrectly predicted, it can have serious consequences, including life
and death. As an example, if malignant tumours are misclassified as benign tumours,
this can pose a great risk to the health of the individual.

10.3.2 Performing at Peak Levels

There is a need to have a high level of performance from a system or model. General
Data Protection Regulation (GDPR) is one example of compliance that should be
explained when compliance is needed.

10.3.3 The Need for Trust

It is essential that the system explains how it generates a particular output when gain‑
ing the trust and confidence of the user. This feature, parameter, and model used in
the application should be identified in the description.

10.4 VARIOUS XAI ALGORITHMS

While there are many XAI algorithms that explain ML models, the following are
three examples:

10.4.1 Shapley Additive Explanations (SHAP)

SHapley Additive exPlanations (SHAP) aims to provide a comprehensive under‑
standing of the technique designed to quantify the contribution of each feature to
a model’s prediction. Through an in‑depth analysis of how predictions shift when
examining all imaginable outcomes, it assigns SHAP values to individual features,
Explainable Artificial Intelligence in the Healthcare 147

thereby allowing straightforward explanations of individual predictions and overall

feature importance [18]. The significant advantage of the SHAP approach has been
extensively explored here, which lies in its ability to offer both local and global inter‑
pretability creating a unified model explanation. This trait makes SHAP an instru‑
mental tool for understanding complex prediction models.

10.4.2 Local Interpretable Model‑Agnostic Explanations

It delves into the application of the Locally Interpretable Model‑Agnostic Explanations
(LIME) for explicating ML models, addressing the challenges of SHAP’s computa‑
tional intensity and time‑extensive nature. LIME enhances these through the genera‑
tion of a data points sample around the predicted data point, followed by constructing
a linear regression model utilizing this weighted sample to discern the impact of
features on prediction [19]. However, it should be noted that LIME’s inability to
maintain global faithfulness is due to its structure that builds explanations solely
based on data point samples adjacent to the instance it explains. The core focus is the
construction and analysis of sparse linear models surrounding individual instances
or predictions done via LIME.

10.4.3 Integrated Gradients
The theory and application of Integrated Gradients is an innovative method for deci‑
phering the predictions of a deep learning (DL) model. The application of Integrated
Gradients involves observing changes in the model’s prediction compared to a base‑
line or masked instance by progressively turning on discrete input features [11,20].
The objective of this method is to identify pivotal features and examine their influ‑
ence on the predictions made by the model. Integrated Gradients offer an optimized
approach by delivering faster computations than SHAP values, holding particular
suitability for DL models. However, this procedure requires the model to possess dif‑
ferentiability for successful implementation. Future research should aim to expand
the applicability of Integrated Gradients to a broader range of models and prediction
tasks [21].

10.5 BEST WORKING ALGORITHM

The short comparative study of two model‑explaining algorithms, LIME and SHAP,
is mentioned in the previous section. The evaluation framework used throughout the
research highlights the strengths and weaknesses of each algorithm, helping readers
to ascertain the best fit based on their requirements and constraints. LIME is identi‑
fied as a preferable choice for scenarios with restricted resources or where immediate
predictions are necessary, due to its primary focus on single data point analysis [22].
In contrast, SHAP emerges as a more efficient tool for those seeking a comprehen‑
sive, global explanation harmonized with specific local predictions. It underscores
the pivotal role of an in‑depth understanding of the underlying data, model intrica‑
cies, and situational nuances in choosing the optimal algorithm.
148 Data-Centric Artificial Intelligence for Multidisciplinary Applications

10.6 EXPLAINABLE AI FOR DECISION‑MAKERS

This delves into the realm of XAI and its significance in rendering ML models
more comprehensible and efficacious for a non‑specialist audience. The research
spotlights the ambiguity in the definitions of interpretability and explainability in
AI literature and explores the increasing appeal of XAI as it presents a solution
to the AI ‘black box’ riddle. It further ventures into the potential uses of AI in the
healthcare sector, particularly highlighting its role in refining treatment regimens
and disease monitoring, while underscoring that most AI models in this field are still
in the nascent stage of development. The document assesses the stakes involved in
complex models and underscores the crucial requirement for transparency and inter‑
pretability, especially when the outcomes are pivotal, as in medical decision‑making
scenarios. It evaluates the different explainability techniques, elaborating on their
respective aims and inherent restrictions. The chapter accentuates the relevance of
XAI frameworks such as LIME, SHAP, and PDP in offering both local and global
insights into the model while stressing the dearth of research on information neces‑
sary for AI‑driven decision‑making in healthcare. The different authors explore the
continual tension between complexity and interpretability in XAI designs, reveal‑
ing the necessity for a multiplicity of explanation methods and a more formalized
approach in this arena [23–26]. This also seeks to illuminate the influence XAI
holds on the conception, acceptance, and utilization of AI‑based instruments in the
healthcare sector.

10.7 NEED OF XAI IN HEALTHCARE

Some use cases within the healthcare industry need to be explained in order to
avoid confusion. The black‑box function of many AI systems is acceptable for most
fields, except the healthcare industry. Many users wish to keep their logic secret and
­private, so they do not want their system to reveal the logic. Black‑box functions,
however, are not acceptable to doctors and patients in healthcare, where mistakes can
lead to dangerous outcomes. In addition to being trained in identifying diseases and
providing treatment, doctors are well‑equipped to treat them. The AI system cannot
diagnose patients if it hasn’t been trained on the correct data. Due to this, users can‑
not be sure of the results of the system, which makes it hard to trust it. In order to
support XAI’s basic principles, transparency, fairness, etc., we must overcompensate
for ML’s opaque nature [27–32].
A system that detects cancer in Caucasian skin is an example of an AI system
that performs better than one that detects cancer in darker‑skinned people. Dark
skin can be missed by this method, thus potentially resulting in cancerous lesions.
This system provides biased results, so they recognize it. Some subpopulations may
be at risk of death as a result of this misjudged output. Insufficient data is later found
to be responsible for this bias. There is not much information about dark skin in
the data set that was used to train the system. Hence, the result, data, and predic‑
tion model of our system need to be more transparent and explained in a clear and
concise manner [33–36].
Explainable Artificial Intelligence in the Healthcare 149

10.8 APPLICATIONS OF XAI IN HEALTHCARE

A great deal of advances in AI have been made in the healthcare industry over the
past few years, and they have proven to be a boon to the industry, including the
detection of genetic links, the use of robotic surgery, and even improving hospital
efficiency due to the use of AI. Figure 10.2 depicts applications of XAI in healthcare.

10.8.1 Clinical Decision Support

A health professional must take into account every crucial detail when diagnosing
a patient. Due to this, medical records often contain unstructured notes that are dif‑
ficult to sift through. If there was even the slightest mistake in monitoring even a
single relevant fact, the life of a patient could be put at risk. In order to narrow down
all relevant information from patient reports, doctors are able to take advantage of

FIGURE 10.2 Applications of XAI in healthcare.

150 Data-Centric Artificial Intelligence for Multidisciplinary Applications

Natural Language Processing (NLP) with the assistance of NLP. With AI, technol‑
ogy can process and store large quantities of information, which will enable database
knowledge to be built and may improve clinical decision support through facilitating
examinations and recommendations for each patient.

10.8.2 Using Chatbots to Improve Primary Care

In many cases, it is common for people to schedule appointments with their gen‑
eral physicians, which may turn out to be a false alarm or a condition that could be
treated by self‑treatment. By automating primary care, AI allows doctors to handle
more crucial and dire cases, without having to worry about mundane administra‑
tive tasks. Using medical chatbots, which are AI‑powered services with smart algo‑
rithms, patients can save money on avoidable doctor’s appointments, thereby saving
time and money on unnecessary doctor’s visits. They provide immediate answers to
health‑related questions and concerns as well as tips on how to handle any potential
problems they may encounter.

10.8.3 Surgeries Using Robots

Incisions made with precision and speed have been revolutionized by AI and col‑
laborative robots. During lengthy and crucial procedures, robots do not get tired, so
fatigue is not an issue. ML can be used to develop new surgical methods using data
from past operations. With these machines, there is little or no chance of uninten‑
tional movements or tremors during surgery. In addition to combining virtual real‑
ity with AI‑enabled robots for minimally invasive surgery, vicarious surgical uses
AI‑enabled robots for minimally invasive surgery, while Carnegie Mellon University
developed Heartlander, a small mobile robot designed to aid in heart therapy.

10.8.4 Nurses on Call
Virtual nursing assistants can direct patients to the best and most effective care units
with the help of AI systems. Almost all queries can be answered by these virtual
nurses around the clock, as well as examinations and instant solutions can be pro‑
vided by them. At the current time, there are many AI‑powered applications that
allow patients to interact with their care providers more regularly between visits to
the doctor’s office, avoiding unnecessary hospital visits by enabling more regular
interactions between patients and their care providers. With AI and voice controls,
Care Angel is the world’s first virtual nurse assistant.

10.8.5 Diagnosing Accurately
With AI, doctors can diagnose diseases more accurately, predict them more effec‑
tively, and diagnose them faster. Additionally, AI algorithms have been proven to be
effective and cost‑effective in diagnosing diabetic retinopathy as well as detecting
other diseases. ML will help pathologists make better diagnoses, for example, thanks
to PathAI. Cancer diagnosis mistakes are being reduced and methods for treating
patients individually are being developed by the company.
Explainable Artificial Intelligence in the Healthcare 151

10.9 AI IN HEALTHCARE: THREATS AND OPPORTUNITIES

It is important to be aware of different risks related to AI in healthcare according to
a report that was released by the Brookings Institution. According to the Institution’s
report, the following threats have been identified as part of the risk management
program all are given in Figure 10.3.

10.9.1 Injury/Error
There is a significant risk that AI in healthcare will at times be wrong, as it might
suggest the wrong treatment to a patient or make the wrong diagnosis on a radiology
scan, resulting in injury or dire health consequences to the patient. For instance, it
might suggest the wrong medication to a patient or make a mistake in locating the
tumour on a radiology scan. At least two reasons may be involved in the difference
between AI errors and human errors. The significance of this is that an AI system
error may cause injury to thousands of patients, while human medical professionals
may also make errors.

10.9.2 Availability of the Data

The fragmented data, coupled with the patient’s tendency to switch insurers and pro‑
viders, causes the data to become more complicated and less understandable, which
increases the possibility of error and increases the total cost.

FIGURE 10.3 Threats of XAI in healthcare.

152 Data-Centric Artificial Intelligence for Multidisciplinary Applications

10.9.3 Security
Many patients believe that data is being collected and exchanged between health
systems and AI developers for the purpose of enabling AI systems, leading them to
sue. AI systems can also predict private information about patients even if they have
not disclosed it themselves. This raises another issue regarding the use of AI systems.

10.9.4 Inequality and Bias

The biases in the available data can also be absorbed by AI systems since they adapt
to the available data. The AI systems are less likely to be aware, and therefore less
likely to treat if the data is primarily collected from academic medical centres.

10.9.5 Changing Professions
Medical professions may undergo significant changes in the future as a result of AI
systems’ use. Most of the work in areas such as radiology is automated. The high
amount of AI use raises the concern that humans will become increasingly unable
to detect AI errors and develop medical knowledge as a result of a decline in human
knowledge and capacity over time.

10.10 HEALTHCARE BENEFITS OF XAI

AI systems in healthcare are rapidly being adopted due to XAI. In the face of Big
Data, a human has difficulty making decisions because AI systems identify pat‑
terns and make decisions based on it. Some features that are provided by XAI are
described in Figure 10.4.

• Transparency: The key to XAI is transparency. User‑friendly algorithms,

models, and features make XAI accessible. The requirements for transpar‑
ency may differ from one user to the next. Suitable users will find it helpful.
• Fidelity: An explanation is provided by the system. The performance of the
model should be in line with what is expected.
• Domain sense: As a result of the system, a user can find an explanation
that is easily understandable to him or her as well as relevant to the domain.
There is a proper context in which it is explained.
• Consistency: In order to avoid confusing the user, explanations should be
consistent for all predictions.
• Generalizability: A general explanation should be provided by the system.
However, it should not be too general.
• Parsimony: There shouldn’t be a lot of complexity in explaining the system.
There should be a maximum amount of simplicity in this process.
• Reasonable: As a result, it explains why AI systems work the way they do.
• Traceable: Data and logic can be tracked by XAI. By knowing how the
data contributed to the output, users are able to make informed decisions.
By tracking logic or data problems, users can solve them in a more logical
or data‑driven manner.
Explainable Artificial Intelligence in the Healthcare 153

FIGURE 10.4 Benefits of XAI in healthcare.

10.11 CHALLENGE OF XAI IN HEALTHCARE

10.11.1 User‑Centric Explanations
Critically examines the challenges involved in creating transparent and explainable
ML models for healthcare systems. It delves into some of the key barriers to this
process, such as the necessity for understanding advanced mathematics and statis‑
tics, inadequate systems design suitable for ML model deployment, the discrepancy
between interpretable explanations and complex ML processes, and difficulties sur‑
rounding the tracking of model training. It also describes the limitations of XAI
methods in identifying relevant features, the essential requirement for a thorough
evaluation of these explanations, and uncertainties regarding the reliability of
XAI methods. It underscores the crucial role of end‑users, specifically clinicians,
in the development and assessment of ML models intended for healthcare settings.
However, it also recognizes the difficulties associated with engaging time‑constrained
154 Data-Centric Artificial Intelligence for Multidisciplinary Applications

clinicians and the discord between the preferences of ML experts and clinicians for
different types of explanations. This advocates effective resolution for these chal‑
lenges to establish more dependable and transparent ML models in healthcare sys‑
tems, thus ensuring their broader acceptability and application.

10.11.2 Performance vs. Transparency Trade‑off

The ramifications of utilizing XAI has been done for enhancing comprehension of
ML models used in healthcare. It discusses the intricate equilibrium between model
complexity and accuracy and elucidates how heightened model transparency can
augment decision analysis. While black‑box AI models such as DL and ensembles
tend to lack transparency and hence inhibit fair decision‑making and trust, grey‑box
AI models, such as statistical models, strike a delicate balance between transparency
and comprehensibility. On the contrary, white‑box AI models like graphical models,
linear models, rule‑based models, and decision trees tend to offer heightened explain‑
ability at the expense of decreased performance. An optimal model is one that offers
high explainability while maintaining satisfactory performance, even in the face of
the inevitable trade‑off of uncovering patterns against attaining data‑fitting accuracy.
The necessity of discussing this trade‑off with end‑users to comprehend the clinical
and human implications of misclassification is also brought to light.

10.11.3 Balancing Requirements of Interpretability

The intricate concept of interpretability in ML systems, specifically within the realm
of healthcare is explored. The focus lies on the quest to decipher a universal definition,
the requisites for an optimal interpretable ML model in healthcare, and the striking
balance between model performance and user comprehension. Delve into the chal‑
lenges encountered in realizing soundness, intelligibility, and scope, alongside instill‑
ing trust regarding the modus operandi of both black and grey box AI models. The
scope of the model is also considered for explanations on both local (instance‑based)
and global levels. Furthermore, how the delicate equilibrium between soundness and
comprehension can be impacted by the specificity of the application domain and the
degree of interpretability anticipated by the end‑user. This seeks to provide a com‑
prehensive understanding of navigational strategies within the complex landscape of
ML system interpretability in healthcare applications.

10.11.4 Assistive Intelligence
The role and limitations of ML algorithms in critical domains, such as healthcare, is
explored. While ML aims to automate decision‑making processes, it is highlighted
that human supervision remains essential, particularly in safety‑critical applica‑
tions. It asserts while ML systems can function as beneficial medical assistants,
they should not be entrusted entirely with patient care due to the need for accu‑
rate data and human intervention. Accordingly, this would lead to arguments for a
human‑in‑the‑loop framework in healthcare and stresses the need for the develop‑
ment of XAI mechanisms to ensure transparency and accountability in ML systems
used in healthcare.
Explainable Artificial Intelligence in the Healthcare 155

10.12 XAI OVER AI: HOW TO OVERCOME CHALLENGES

Customer trust is one of the biggest challenges in AI systems. An opaque AI system
is one in which the system’s output is not accompanied by any explanation or hypoth‑
esis. The healthcare system is an example of where customers find it challenging to
trust machines that do not explain. Medical practitioners and patients do not adopt
opaque AI systems due to their incompetence. This is reduced by XAI.

10.12.1 Trust and Confidence

In the AI system, doctors and patients can become wary of it because it is opaque.
For various reasons, users seek explanations of the system in order to gain a better
understanding of the logic of the model, as well as to give an explanation to others
as to why the system works and to provide the reasoning behind making a decision.
XAI builds users’ trust and confidence by providing them with explanations.

10.12.2 Detect and Remove Bias

Because of its lack of transparency, the system cannot provide users with a clear view
of its defects and biases. As a result, the detection and removal of bias and the provi‑
sion of safeguards against bias become difficult.

10.12.3 Model Performance
The model’s behaviour cannot be tracked due to the lack of awareness among model
users.

10.12.4 Regulatory Standards
Regulatory standards cannot be recognized by users. As a result, the system would
be harmed.

10.12.5 Risk and Vulnerability

Risk management systems must be able to explain how they deal with risks.
Particularly, when the environment is unknown to the user, it can be detected early
and action can be taken with XAI. What if the system is unable to provide the user
with information about how they can mitigate the risks associated with it?

10.13 FUTURE PROSPECTS

Models can be used to simplify complex systems that are intractable. ML capabili‑
ties may be underutilized if DNNs with many parameters are not exploited in order
to extract quality data using DNNs. A new, more complex component can be added
to an existing model with the intent of gaining new insights. Just as it was with the
earlier models and demonstrations, it’s very important to discuss why the additional
156 Data-Centric Artificial Intelligence for Multidisciplinary Applications

components correlate with other insights. In some industries, there are still a number
of techniques that are embedded deeply into the culture, yet powerful ML algorithms
are finding new applications in these fields. Fragmented and experimental implemen‑
tations of existing or custom‑developed interpretable techniques exist in the nascent
field of medical ML, which is still in its infancy. In spite of the current focus on
improving feature selection and extraction accuracy and performance, interpretabil‑
ity research may still have large untapped potential.

10.14 CONCLUSION
This chapter has significantly explored the profound relevance of XAI in healthcare,
underlining its critical role in high‑stakes decision‑making processes that pervade
the medical field. In light of potentially severe consequences arising from incor‑
rect predictions by AI models, this underscores the need for the development and
adoption of techniques for building AI applications that aid users in comprehending
the model’s output and predictions. The chapter detailed various methods, assign‑
ing them into six categories, and thoroughly examined diverse tools and techniques
specific to healthcare’s unique demands. Moreover, the chapter sheds light on the
irrefutable demand for integrated explainability tools combining XAI, ML, and DL,
especially for delicate medical procedures that call for utmost precision. Recognizing
the increasing demand for trustworthy and transparent AI models among medical
professionals, the call for enhanced collaboration between data scientists and medi‑
cal experts for the design and successful development of efficient XAI systems is
indispensable. Successfully harnessing this alliance can lead to an enhanced under‑
standing of diseases’ causation, better measurement of medication influences, and
overall improved patient satisfaction. Finally, the chapter excelled in providing fur‑
ther research sources, reinforcing its arguments and providing groundwork for future
exploration in the realm of XAI in healthcare. Such ingenuity creates an opportunity
to delve deeper into the concept and critically analyze areas such as ethics and the
engagement of user‑centred design in AI developments. This ultimately paves the
way for the potential maximization of the use of AI, particularly XAI in healthcare,
advancing medical practices towards remarkable progress and efficiency.

REFERENCES
1. Arrieta AB, Díaz‑Rodríguez N, Del Ser J, Bennetot A, Tabik S, Barbado A et al.
Explainable artificial intelligence (XAI): Concepts, taxonomies, opportunities and
challenges toward responsible AI. Information Fusion. 2020;58:82–115.
2. Angelov PP, Soares EA, Jiang R, Arnold NI, Atkinson PM. Explainable artificial
intelligence: An analytical review. Wiley Interdisciplinary Reviews: Data Mining and
Knowledge Discovery. 2021;11(5):e1424.
3. Samek W, Wiegand T, Müller K‑R. Towards explainable artificial intelligence,
Explainable AI: Interpreting, Explaining and Visualizing Deep Learning, pp. 5–22.
doi:10.1007/978‑3‑030‑28954‑6_1. 2017.
4. Ribeiro MT, Singh S, Guestrin C, editors. Why should i trust you? Explaining the pre‑
dictions of any classifier. In: Proceedings of the 22nd ACM SIGKDD international
Conference on Knowledge Discovery and Data Mining, 2016. San Francisco, USA.
Explainable Artificial Intelligence in the Healthcare 157

5. Zhou B, Khosla A, Lapedriza A, Oliva A, Torralba A, editors. Learning deep features

for discriminative localization. In: Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition, 2016. Las Vegas, NV, USA.
6. Selvaraju RR, Cogswell M, Das A, Vedantam R, Parikh D, Batra D, editors. Grad‑cam:
Visual explanations from deep networks via gradient‑based localization. In: Proceedings
of the IEEE International Conference on Computer Vision, 2017. Venice, Italy.
7. Chattopadhay A, Sarkar A, Howlader P, Balasubramanian VN, editors. Grad‑cam++:
Generalized gradient‑based visual explanations for deep convolutional networks. In:
2018 IEEE Winter Conference on Applications of Computer Vision (WACV), IEEE,
2018. Lake Tahoe, NV, USA.
8. Cirqueira D, Nedbal D, Helfert M, Bezbradica M, editors. Scenario‑based require‑
ments elicitation for user‑centric explainable AI: A case in fraud detection. In: Machine
Learning and Knowledge Extraction: 4th IFIP TC 5, TC 12, WG 84, WG 89, WG 129
International Cross‑Domain Conference, CD‑MAKE 2020, Proceedings 4, Dublin,
Ireland, August 25–28, 2020, Springer, 2020.
9. Alicioglu G, Sun B. A survey of visual analytics for explainable artificial intelligence
methods. Computers & Graphics. 2022;102:502–20.
10. Kim B, Wattenberg M, Gilmer J, Cai C, Wexler J, Viegas F, editors. Interpretability
beyond feature attribution: Quantitative testing with concept activation vectors (tcav).
In: International Conference on Machine Learning, PMLR, 2018. Stockholm Sweden.
11. Raghunath MP, Ankayarkanni B. An empirical review on brain tumor classification
approaches. Information Systems and Management Science, 2022;12:15–27.
12. Noh H, Hong S, Han B, editors. Learning deconvolution network for semantic seg‑
mentation. In: Proceedings of the IEEE International Conference on Computer Vision,
2015., pp. 114–123, vol‑07.
13. Watson D. The rhetoric and reality of anthropomorphism in artificial intelligence.
Minds and Machines. 2019;29(3):417–40.
14. Guidotti R, Monreale A, Ruggieri S, Turini F, Giannotti F, Pedreschi D. A survey of meth‑
ods for explaining black box models. ACM Computing Surveys (CSUR). 2018; 51(5):1–42.
15. Misztal‑Radecka J, Indurkhya B, editors. Getting to know your neighbors (KYN).
Explaining item similarity in nearest neighbors collaborative filtering recommen‑
dations. In: Adjunct Publication of the 28th ACM Conference on User Modeling,
Adaptation and Personalization, 2020. Italy.
16. Thommandru, A., Mutkule, P., Bandi, A., and Tongkachok, K. Towards applicability of
artificial intelligence in healthcare, banking and education sector. ECS Transactions.
2022;107(1):16665–16671.
17. Nimmy SF, Hussain OK, Chakrabortty RK, Hussain FK, Saberi M. Explainability
in supply chain operational risk management: A systematic literature review.
Knowledge‑Based Systems. 2022;235:107587.
18. Saeed W, Omlin C. Explainable ai (xai): A systematic meta‑survey of current chal‑
lenges and future opportunities. Knowledge‑Based Systems. 2023;110273.
19. Egger J, Pepe A, Gsaxner C, Jin Y, Li J, Kern R. Deep learning‑a first meta‑survey
of selected reviews across scientific disciplines, their commonalities, challenges and
research impact. PeerJ Computer Science. 2021;7:e773. 56.
20 Gritzalis D, Iseppi G, Mylonas A, Stavrou V. Exiting the risk assessment maze: A
meta‑survey. ACM Computing Surveys (CSUR). 2018;51(1):1–30.
21. Došilović FK, Brčić M, Hlupić N, editors. Explainable artificial intelligence: A sur‑
vey. In: 2018 41st International Convention on Information and Communication
Technology, Electronics and Microelectronics (MIPRO), IEEE, 2018.
22. Kolli CS, Raghunath MP, Meenakshi S, Maheswari K, Britto CF, Kushwaha S, 2022.
Efficient development of supervised learning algorithm for kidney stone prediction. In:
2022 International Conference on Inventive Computation Technologies (ICICT). Nepal.
158 Data-Centric Artificial Intelligence for Multidisciplinary Applications

23. Das A, Rad P. Opportunities and challenges in explainable artificial intelligence (xai):
A survey (2020) pp. 229–239, vol‑11.
24. Alonso JM, Catala A, editors. Proceedings of the 1st Workshop on Interactive
Natural Language Technology for Explainable Artificial Intelligence (NL4XAI 2019).
Proceedings of the 1st Workshop on Interactive Natural Language Technology for
Explainable Artificial Intelligence (NL4XAI 2019) (2019). Turkey.
25. Almasri A, Alkhawaldeh RS, Çelebi E. Clustering‑based EMT model for predicting
student performance. Arabian Journal for Science and Engineering. 2020;45:10067–78.
26. Warman A, Warman PI, Sharma A, Parikh P, Warman R, Viswanadhan N et al.
Interpretable artificial intelligence for COVID‑19 diagnosis from chest CT reveals
specificity of ground‑glass opacities (2020). doi:10.1101/2020.05.16.20103408.
27. Zhu S, Fan W, Yang S, Pardalos PM. Scheduling operating rooms of multiple hospi‑
tals considering transportation and deterioration in mass‑casualty incidents. Annals of
Operations Research. 2023;321(1‑2):717–53.
28. Lan S, Fan W, Shao K, Yang S, Pardalos PM, editors. Medical Staff Scheduling
Problem in Chinese Mobile Cabin Hospitals During Covid‑19 Outbreak. Learning and
Intelligent Optimization: 15th International Conference, LION 15, Athens, Greece,
June 20–25, 2021, Revised Selected Papers 15, Springer, 2021.
29. Lan S, Fan W, Yang S, Pardalos PM, Mladenovic N. A survey on the applications of vari‑
able neighborhood search algorithm in healthcare management. Annals of Mathematics
and Artificial Intelligence. 2021;12:1–35.
30. Anand L, Rane, K.P., Bewoor, L.A., Bangare, J.L., Surve, J., Raghunath, M.P.,
Sankaran, K.S., and Osei, B., 2022. Development of machine learning and medical
enabled multimodal for segmentation and classification of brain tumor using MRI
images. Computational Intelligence and Neuroscience, 2022;29:1–8.
31. Fan W, Liu J, Zhu S, Pardalos PM. Investigating the impacting factors for the healthcare
professionals to adopt artificial intelligence‑based medical diagnosis support system
(AIMDSS). Annals of Operations Research. 2020;294:567– 92.
32. Abidin, S., Prasad Raghunath, M., Rajasekar, P., Kumar, A., Ghosal, D., and Ishrat,
M., 2022. Identification of disease based on symptoms by employing ML. In: 2022
International Conference on Inventive Computation Technologies, Nepal (ICICT).
33. Horáček J, Koucký V, Hladík M. Novel approach to computerized breath detection in
lung function diagnostics. Computers in Biology and Medicine. 2018;101:1–6.
34. Pardalos, P. M., Georgiev, P. G., Papajorgji, P., and Neugaard, B. (eds.). Systems analy‑
sis tools for better health care delivery. Springer Science & Business Media, 2013;
vol. 74.
35. Alves, C.J., Pardalos, P.M., and Vicente, L.N. (Eds.). Optimization in medicine. Springer
Science & Business Media, 2007; vol. 12.
36 Roshanzamir, M. et al. (2023) Quantifying uncertainty in automated detection of
alzheimer’s patients using Deep Neural Network [Preprint]. doi:10.20944/preprints
202301.0148.v1.
 11 Role of Data‑Centric
Artificial Intelligence
in Agriculture
Rajkumar Patil, Nilesh Popat Sable,
Parikshit N. Mahalle, Gitanjali R. Shinde,
Prashant Dhotre, and Pankaj Chandre

11.1 INTRODUCTION
Indian agriculture has traversed a remarkable journey from an era where farmers
relied on age‑old techniques and manual labor to the present, characterized by the
integration of modern tools and technology. In the past, agricultural practices were
deeply rooted in traditional wisdom, with manual plowing, hand sowing, and rudi‑
mentary irrigation methods being the norm. As time progressed, the Green Revolution
introduced improved seeds, fertilizers, and mechanized equipment, catapulting pro‑
ductivity. Today, precision agriculture, satellite imagery, IoT devices, and data‑driven
insights have reshaped Indian farming, enhancing efficiency and sustainability while
bridging the gap between historical practices and cutting‑edge innovation [1].
The realm of agriculture is undergoing a transformative evolution, powered by
the synergistic integration of Data‑Centric Artificial Intelligence (AI) [2–4]. This
convergence holds the promise of addressing pressing challenges in food production,
resource allocation, and sustainability. With an ever‑growing global population and
the escalating impact of climate change, the need for innovative solutions in agricul‑
ture has never been more critical.
The advent of AI has revolutionized the way we perceive and harness data.
Agriculture, a sector deeply rooted in empirical knowledge and practice, is now
embracing the data‑driven paradigm. Data, often referred to as the new “oil,” has
become a valuable asset for decision‑makers across the agricultural spectrum. By
capturing and analyzing a wealth of information ranging from climate patterns, soil
health, crop growth, and market trends, AI empowers stakeholders with insights
that were once beyond reach [5]. The introduction of AI in agriculture is not merely
a technological shift, but a strategic shift in the very fabric of farming practices.
Traditional approaches, while valuable, often grapple with inefficiencies and uncer‑
tainties that can limit productivity. In contrast, Data‑Centric AI offers the potential
to enhance precision, optimize resource utilization, and minimize environmental
impact. Through advanced predictive models, decision support systems (DSSs), and

DOI: 10.1201/9781003461500-13 159

160 Data-Centric Artificial Intelligence for Multidisciplinary Applications

smart automation, AI‑enabled agriculture holds the promise of achieving higher

yields while conserving resources and minimizing waste [6].
The objective of this chapter is to examine the crucial part that data‑centric AI
plays in agriculture. It delves into the various dimensions of this transformation,
starting with the convergence of AI and Agriculture, highlighting the paramount
importance of data in modern agriculture. After that, it examines several approaches
of gathering data in agriculture, including sensor technology, Internet of Things
(IoT) applications, remote sensing, satellite images, and aerial surveillance. Further
exploration of the complex procedures for data preparation, cleaning, and applica‑
tion of machine learning (ML) methods – particularly deep learning (DL) for image
analysis – is provided in this chapter. The chapter next looks at the use of predictive
modeling in agriculture, discussing forecasts for crop production, pest and disease
outbreaks, and the effect of weather on agriculture. The adoption of smart agricul‑
tural equipment, variable rate technology, precise irrigation and fertilization, and
resource optimization are all covered in‑depth. The chapter also looks at the creation
of warnings, recommender systems for agricultural practices, risk assessment, and
mitigation, as well as DSSs for real‑time monitoring. Lastly, it acknowledges the
challenges and future directions in Data‑Centric AI in agriculture, touching upon
concerns related to data privacy and security, integration with traditional farming
knowledge, and the scalability of AI solutions for smallholder farmers. The chapter
concludes with a recap of the benefits and opportunities presented by Data‑Centric
AI in agriculture, accompanied by a call to action for future development in this field.

11.1.1 The Convergence of AI and Agriculture

The convergence of AI and agriculture is an innovative synergy that has the potential
to transform how we produce food, manage resources, and tackle the urgent problems
of feeding a growing global population while minimizing environmental impacts.
Agriculture is being improved in every way by the use of AI technologies like ML,
computer vision, and data analytics. In precision agriculture, large datasets from
sensors, satellites, and drones are analyzed by AI‑driven algorithms to improve crop
management [7]. In order to maximize yields and optimize resource use, farmers can
make data‑driven decisions about planting, irrigation, and pest control [8]. Robotic
systems that are AI‑powered are also creating a name for themselves by automating
labor‑intensive operations like weeding and harvesting, lowering labor costs, and
boosting output [9]. Additionally, AI‑driven models which predict weather patterns
and crop diseases give farmers the ability to plan and act in advance, eventually
protecting their livelihoods [10]. In this era of convergence, AI is not only increasing
output but also supporting sustainability by lowering chemical inputs, conserving
water, and reducing waste. It’s promoting a new era of “smart farming,” in which
automated systems and data‑driven insights enable farmers to adapt to shifting envi‑
ronmental circumstances and maximize resource utilization. However, issues still
exist, such as ensuring that small‑scale farmers have equitable access to AI tools and
resolving data privacy issues. Despite this, the combination of AI and agriculture has
great potential for creating food production systems that are more effective, robust,
Role of Data‑Centric Artificial Intelligence in Agriculture 161

and sustainable while using fewer resources to fulfill the needs of the world’s rising
population. Figure 11.1 shows the potential application of AI in Agriculture.
The coronavirus epidemic and the Ukraine War, which were both exacerbated
by a labor shortage, had a negative impact on the record levels of crop production in
2022 [11]. According to the researcher’s findings in Ref. [12], the widespread adop‑
tion of AI and Precision Agriculture tools has the potential to significantly lower
operating expenses as a percentage of revenue, from 42% to 33%. A potential $67
billion market opportunity could arise from this. Figure 11.2 shows global annual
operating cost compared with the use of AI for the year 2022 for Corn, Wheat, and
Soybean. Also, when autonomous technology becomes more widely used in agricul‑
ture, businesses may be able to create recurring revenue streams with margins com‑
parable to software‑as‑a‑service business models. It’s crucial to remember that, even

FIGURE 11.1 Application of AI in agriculture.

162 Data-Centric Artificial Intelligence for Multidisciplinary Applications

FIGURE 11.2 Global annual agricultural operating cost. Source: ARK Investment
Management LLC, 2023, based on data from USDA as of July 12, 2023 [12].

though the emphasis has been on cost reduction, autonomous solutions can increase
crop yields by reducing the need for human labor, particularly at night or during cru‑
cial agronomic periods. Due to labor constraints, many farmers sometimes struggle
to procure labor‑dependent machinery for field activities. Autonomous technologies
have the potential to dramatically improve crop output prospects while addressing
this difficulty.
We hold the belief that AI and Precision Agriculture have the potential to bring
about the most significant advancements in agriculture since the introduction of the
tractor a century ago. In our perspective, these innovations have the capacity to boost
farm profitability, reduce food costs, and meet the growing global need for crops,
ultimately enhancing the efficiency and sustainability of farming on a global scale
within the next 5–10 years.

11.1.2 Importance of Data in Modern Agriculture

Data is essential to modern agriculture and has transformed the sector in several ways.
Farmers can now gather and analyze a ton of data from sensors, drones, and satel‑
lites due to precision agriculture in order to make wise decisions. This data‑driven
strategy enables exact resource allocation, optimizing the use of pesticides, fertil‑
izers, and water, while minimizing waste and adverse environmental effects. Data
also helps with better pest management, disease outbreak prediction, and crop health
monitoring. It improves logistics along the supply chain, assuring prompt deliveries
Role of Data‑Centric Artificial Intelligence in Agriculture 163

and cutting down on food waste. In the end, data‑driven agriculture promotes sus‑
tainability, boosts yields, and aids in feeding a growing world population while
minimizing negative environmental effects. In Ref. [13], the author highlights the
pivotal role of efficient data management in driving the exponential growth of Smart
Farming in modern agriculture. It emphasizes the use of data in supporting pro‑
ducers’ crucial decisions, with a focus on maximizing productivity and sustainabil‑
ity through unbiased data gathered from sensors. It has been demonstrated that the
future of sustainable agriculture will be built on the integration of AI and data‑driven
tactics with robotic solutions. The review thoroughly examines the entire range of
advanced farm management systems, from data collection in crop fields to variable
rate applications, highlighting their potential to improve resource utilization, lessen
environmental impact, and revolutionize food production to address the challenges
of impending population growth.
The primary data collected from crops must be processed efficiently to transform
numerous images into meaningful and clear‑cut information. Farmers in traditional
settings rely on visual inspections to make crop management decisions based on their
experience without the use of technological technology. Farms with cutting‑edge
technology tend to adopt a more data‑driven strategy. Sensors are used to gather fac‑
tual information about the environment, the soil, and the crops. To help farmers make
wise decisions, the data is subsequently processed using AI algorithms and filtering
techniques. Throughout the information‑based management cycle for advanced agri‑
culture shown in Figure 11.3, this cycle – from data gathering to action – continues
and comes to an end after harvest. The way that farmer maximizes their crop yields
and resource management has been revolutionized by precision agriculture.

11.2 DATA COLLECTION IN AGRICULTURE

Modern farming practices rely heavily on data collecting that helps farmers and
researchers make educated decisions, and increase productivity while minimizing
resource wastage. Data on soil, weather, crop growth, and livestock health are all
systematically gathered as part of this process. The ideal circumstances for planting
and fertilizing crops can be determined using information about the soil, including
its pH, nutrient content, and moisture levels. The ability to predict weather patterns
and make prompt decisions about irrigation and pest management is made possible
by weather data such as temperature, precipitation, and humidity. Sensor and drone
data on crop growth are used to track the health of plants, their rates of expansion,
and any possible problems like disease or nutritional deficiencies. The tracking of
an animal’s weight, health, and reproductive habits is included in livestock statistics.
Data collecting in agriculture has improved in accuracy and efficiency with the intro‑
duction of cutting‑edge technology like IoT sensors, remote sensing, and ML. These
data‑driven insights help farmers use resources more efficiently, lessen their impact
on the environment, and improve yields, all of which improve the viability and prof‑
itability of agricultural practices. Additionally, c data sharing and collaboration
within the agricultural community are essential to harness the full potential of data
collection, allowing farmers to adapt to changing conditions and global challenges
like climate change and food security. In summary, agricultural data collecting has
164 Data-Centric Artificial Intelligence for Multidisciplinary Applications

FIGURE 11.3 Information‑based management cycle for advanced agriculture.

transformed into a crucial instrument for modern farming, fostering innovation and
efficiency in a sector entrusted with feeding a growing global population while con‑
tending with a variety of environmental and economic concerns.

11.2.1 Internet of Things: Collecting Information

In the realm of agriculture, the Internet of Things (IoT) involves the deployment of
sensors and various devices to transform every aspect and activity of farming into
valuable data. Approximately 10%–15% of farmers in the United States are report‑
edly utilizing IoT solutions on their farms, covering an impressive 1.2 billion ha and
encompassing 250,000 farms [14]. IoT plays a pivotal role in Agriculture 4.0, and
it’s a driving force behind the collection of significant data in agriculture [15]. The
agricultural sector is poised to undergo substantial transformation owing to advance‑
ments in IoT technologies [16]. Projections indicate that, with the adoption of new
techniques, IoT has the potential to boost agricultural productivity by a remarkable
Role of Data‑Centric Artificial Intelligence in Agriculture 165

70% by the year 2050 [17]. This is particularly crucial as the world faces the chal‑
lenge of increasing global food production by 60% by 2050 to support a growing
population expected to reach over nine billion [18].
One of the primary benefits of IoT implementation is the achievement of higher
crop yields and cost reduction. For instance, studies conducted by OnFarm revealed
that, on average, farms using IoT experience a 1.75% increase in crop yields, while
energy costs decrease by $17–$32 per ha, and the water usage for irrigation dimin‑
ishes by 8% [15].

A. Sensor Technologies
Sensor technologies have played a significant role in revolutionizing agri‑
culture by enabling farmers to monitor and manage their crops, livestock,
and overall farm operations more efficiently. These sensors gather data on
many environmental aspects and offer insightful analysis to enhance deci‑
sion‑making and optimize resource allocation [19]. In agriculture, the fol‑
lowing prominent sensor technologies are employed:
1. Soil Sensors
a. Soil Moisture Sensors: Let farmers decide when and how much to
irrigate crops, therefore saving water and enhancing crop develop‑
ment. These sensors assess the moisture content in the soil.
b. Soil pH Sensors: Measure the acidity or alkalinity of the soil, which
is important for managing nutrients and choosing the right crops.
2. Weather Stations
a. Weather Stations: Include numerous sensors including tempera‑
ture, humidity, wind speed, and rainfall detectors. Weather stations
also incorporate other types of sensors. For accurate weather fore‑
casts and administration of farming activities, they offer real‑time
weather data.
3. Crop Health Sensors
a. Remote Sensing: To track crop health, find illnesses, and determine
nutrient levels, satellite‑ or drone‑based sensors collect photos and
data.
b. Hyper spectral Imaging: This cutting‑edge technique examines the
light reflected from crops to detect minute changes in plant health.
4. Livestock Monitoring Sensors
a. RFID Tags and GPS: Used in the location and movement monitor‑
ing of cattle. Information on breeding and health can also be stored
on RFID tags.
b. Wearable Sensors: To keep checks on an animal’s health, behavior,
and well‑being, sensors like accelerometers and temperature gauges
can be fastened to it.
5. Environmental Sensors
a. Air Quality Sensors: Measure factors such as air temperature,
humidity, and gas concentrations to ensure optimal conditions for
livestock and crops.
166 Data-Centric Artificial Intelligence for Multidisciplinary Applications

b. Water Quality Sensors: Check the purity of water sources before

using them for irrigation or animal use.
6. Nutrient Management Sensors
a. Nitrate and Phosphate Sensors: Enable farmers apply fertilizer
accurately to avoid excess or underuse by measuring nutrient levels
in soil or water.
7. Pest and Disease Monitoring
a. Pest Traps and Sensors: These tools enable focused interventions by
assisting in the detection and monitoring of pest and disease activity
in crops.
B. Drones and Aerial Surveillance
Aerial surveillance and drone technology have revolutionized agriculture
by allowing for more precise and effective crop management. The high‑res‑
olution cameras and a variety of sensors on board these unmanned aerial
vehicles give farmers crucial information for streamlining their operations.
Drones give farmers the ability to make data‑driven choices about anything
from crop health monitoring and insect infestation detection to determin‑
ing irrigation requirements and yield estimates. Aerial surveillance helps
in early problem identification, enables prompt response, decreases crop
losses, and reduces the need for pesticides and water resources [20]. Drones
are also useful for mapping fields and producing in‑depth 3D models, which
helps with land management and boosts overall production. Drones and
aerial surveillance are turning into increasingly important tools for modern
farmers as they strive to increase yields, cut costs, and promote environ‑
mentally friendly farming practices as the agricultural sector continues to
face challenges like climate change and the need for sustainable practices.
In Figure 11.4, the integration of sensor technologies, drones, and aerial
surveillance exemplifies their pivotal roles in IoT‑driven smart agriculture.
By gathering data on crop health, weather, and soil moisture in real time,

FIGURE 11.4 IoT‑based smart agriculture. Source: From ref. [20].

Role of Data‑Centric Artificial Intelligence in Agriculture 167

these technologies improve precision farming and help farmers manage

resources more effectively and produce more food.

11.3 DATA PROCESSING AND ANALYSIS

Data processing and analysis in agriculture play a pivotal role in modern farming
practices, contributing to increased productivity, resource optimization, and informed
decision‑making discussed in Section 11.1 in detail. Farmers and agricultural profes‑
sionals utilize various data sources, including weather data, soil quality measure‑
ments, crop yields, and remote sensing imagery, to gather valuable information about
their fields discussed in Section 11.2 in detail. Through advanced technologies such
as IoT sensors, drones, and satellite imaging, these data are collected in real time
or at regular intervals discussed in Section 11.2.1 in detail. Subsequently, this raw
data is processed using specialized software and algorithms to extract meaningful
insights. Analysis of this processed data helps farmers make informed choices about
crop planting, irrigation, fertilizer application, and pest control. It also enables the
prediction of disease outbreaks and adverse weather events, thus allowing for proac‑
tive measures. Ultimately, data processing and analysis empower agriculture with
data‑driven precision, leading to more sustainable and efficient farming practices
while ensuring environmental conservation.

11.3.1 Data Preprocessing
Preprocessing the data is an essential stage in agricultural data analysis because it
guarantees that the information that will be used for decision‑making and analysis is
correct, dependable, and consistent. Figure 11.5 illustrates some key steps and factors
involved in preprocessing of data for agricultural purposes.

Integration of Data: To build a single dataset, combine information from sev‑

eral sources, such as weather, soil, and crop yield data. Ensure that the data
is consistent with respect to time, place, and other important properties [21].
Transformation of Data: To bring all variables to a common scale, we normal‑
ize or scale the data. If distributions of data are skewed, use transformations
like logarithms to handle them. Use of one‑hot encoding to convert category
variables into numerical representation [22].
Feature Engineering: Improve the performance of the model by developing
additional features that are based on domain knowledge. Extract the infor‑
mation that is pertinent from the raw data, such as generating vegetation
indices from satellite images or heat stress indices [23].
Dimensionality Reduction: Reduce the amount of features in order to avoid the
“curse of dimensionality” and boost the overall efficiency of the model. In
order to accomplish this goal, one may make use of methods such as princi‑
pal component analysis [24].
Data Cleaning: Deal with missing data by either assuming their values or inter‑
polating between them. Eliminate any outliers that can bias the outcomes of
the study or modeling. Find and fix any inaccuracies in the measurements
or data entries [25].
168 Data-Centric Artificial Intelligence for Multidisciplinary Applications

FIGURE 11.5 Key steps for data preprocessing.

The process of preprocessing data is iterative, and the choice of preprocessing pro‑
cesses will rely on the particular agricultural application as well as the features
of the data. An efficient preprocessing of data can result in better predictions and
insights, which can, in turn, lead to improvements in agricultural decision‑mak‑
ing, whether the prediction is for crop production, disease detection, or resource
allocation.

11.3.2 ML and DL Techniques in Agriculture

The management of crops, the prediction of yields, and the efficient use of resources
have all been transformed by ML and DL. By providing farmers with data‑driven
insights, these innovative solutions are promoting sustainable and effective agri‑
cultural practices. ML algorithms like Random Forests, Support Vector Machines,
and Decision Trees are invaluable tools for tasks such as disease detection, pest
monitoring, and yield prediction. As a result, farmers are able to proactively address
problems and improve their agricultural operations. They do this by using past data
Role of Data‑Centric Artificial Intelligence in Agriculture 169

to find trends and make predictions. For instance, ML models can assist in early
disease detection, enabling prompt action, and minimizing agricultural losses by
analyzing data on crop health and environmental conditions. DL, with its neural
network architectures, has made significant strides in agriculture. Convolutional
neural networks (CNNs) are excellent in image analysis, which makes them the best
choice for identifying plant diseases. CNNs can quickly and effectively identify
illnesses by studying images of leaves or fruits, eliminating the need for manual
checks and resulting in healthier crops [26]. Figure 11.6 shows the architecture of
ML and DL in agriculture.
Another DL method used in agriculture for time series data analyzing, such as
weather forecasting, is recurrent neural networks (RNNs). RNN can analyze huge
amounts of historical meteorological data to predict future circumstances, assisting
farmers in making well‑informed choices about when to sow, irrigate, and harvest.
This enhances resource allocation and minimizes the impact of weather‑related
risks on crop yields [27]. Generative adversarial networks (GANs) play a unique role
by generating synthetic data to simulate various environmental conditions. Without
the requirement for actual field testing, this synthetic data helps with crop plan‑
ning and scenario testing. For optimal production and resource efficiency, farmers
can utilize GANs to optimize resource allocation, irrigation plans, and crop rota‑
tion techniques [28]. Figure 11.7 illustrates the agricultural challenges addressed by
various research projects, along with the AI‑powered solutions employed for each of
these endeavors. These projects differ in their goals, approaches, and the materials
utilized [29].
Due to the tremendous capabilities that ML and DL technologies have given
farmers for decision‑making and resource optimization, agriculture has undergone a
revolution. These innovations improve weather forecasting, disease detection, yield
prediction, and crop management. Farmers may contribute to sustainable agricul‑
tural practices that are essential for feeding a growing world population by utiliz‑
ing the potential of ML and DL, which can also help them enhance production and
minimize resource waste.
Figure 11.7 shows the Conceptualization of the several agricultural operations and
AI‑related technologies of the European research projects.

FIGURE 11.6 Basic architecture of ML and DL for agriculture.

170 Data-Centric Artificial Intelligence for Multidisciplinary Applications

FIGURE 11.7 Conceptualization of the several agricultural operations and AI‑related tech‑
nologies of the European research projects described in [29].

11.4 DECISION SUPPORT SYSTEMS

The integration of Data‑Centric AI and DSSs has become a transformational force in
the field of modern agriculture. By offering real‑time monitoring and alerts, enabling
effective recommender systems for agricultural practices, and boosting risk assess‑
ment and mitigation tactics, these technologies provide farmers with invaluable tools
[30]. The development of sensor technology and remote monitoring tools has made
it possible for DSS to continually gather and analyze information about crop health,
soil moisture, weather patterns, and equipment status. Farmers can now make quick
decisions about irrigation, pest management, and resource allocation. Additionally,
DSSs may notify farmers via alerts and notifications of prospective problems like
unfavorable weather patterns or disease outbreaks, allowing for prompt measures
to safeguard crops and increase yields. Figure 11.8 presents two categories of DSS
the initial category enables data representation through visualizations like maps or
graphs, while the second category focuses on forecasting future events (predictive).
In the second phase that Data‑centric AI plays a crucial role, they possess the capac‑
ity to glean insights from extensive historical data [30].
AI‑powered recommender systems are essential for improving farming methods.
These systems can provide specific suggestions to farmers by compiling information
on soil quality, previous crop performance, market trends, and weather forecasts.
This might involve suggestions for planting seasons, crop rotation, and even crop
price predictions from the market. Farmers may maximize resource use and raise
overall production by implementing these tips.
Risk assessment and mitigation are critical components of DSS in agriculture. AI
systems can examine historical data to find trends in crop diseases, insect infesta‑
tions, or extreme weather occurrences. Armed with this information, farmers may
Role of Data‑Centric Artificial Intelligence in Agriculture 171

FIGURE 11.8 Architecture of DSS and data‑centric AI for smart agriculture.

proactively employ risk mitigation techniques, such as modifying planting dates,

selecting crop types that are disease‑resistant, or spending money on protective
infrastructure. In addition to reducing potential losses, this proactive strategy sup‑
ports resilient and sustainable agricultural methods [31,32].
Data‑Centric AI in agriculture, supported by DSSs, has revolutionized the indus‑
try. These technologies provide real‑time data, personalized advice, and sophisti‑
cated risk assessment capabilities to farmers. To ensure food security for a growing
global population while reducing environmental impact, agribusiness may advance
toward more effective, lucrative, and sustainable practices by utilizing the potential
of AI‑driven DSS.

11.5 CHALLENGES AND FUTURE DIRECTIONS

As agriculture transitions to the digital era, data‑centric AI has enormous potential to
transform the sector. To realize its full potential, a number of significant issues and
future initiatives must be addressed.

11.5.1 Data Privacy and Security Concerns

The biggest difficulty faced in data‑centric AI for agriculture is how to protect the
growing troves of sensitive data. In the modern farming environment, a wide range
of IoT devices, drones, and satellite technologies come together to collect intricate
data about crops and soil conditions. However, the influx of such priceless data raises
the unsettling possibility of privacy invasions, exposing the susceptibility of per‑
sonal agricultural information to malicious actors. A determined effort is required
to address this critical problem, including the development and implementation of
powerful data encryption, access control, and anonymization solutions. These tech‑
nological fortifications will serve as an indomitable bulwark against the perils of data
exposure. Furthermore, ensuring data privacy and security across the agriculture
sector depends on the development of thorough rules and industry standards. We
can protect farmers’ interests while ensuring the continued development and innova‑
tion of data‑centric AI in agriculture by creating a strong framework of rules and
standards. We may strive to leverage the transformational potential of data‑driven
172 Data-Centric Artificial Intelligence for Multidisciplinary Applications

agriculture while protecting the sensitive data that drives its development in this
harmonic synergy of technology, legislation, and monitoring [33].

11.5.2 Integration with Traditional Farming Knowledge

The integration of AI and data‑driven technology offers intriguing potential in
the constantly changing world of agriculture, but it is crucial that these develop‑
ments live in harmony with the research of conventional agricultural knowledge.
The invaluable practical insights and expertise of generations of farmers must not
be relegated to the periphery. Instead of replacing these time‑honored talents, AI
should be seen as a potent ally that enhances and perfects them. The establishment
of user‑friendly AI systems that are meticulously developed to promote interac‑
tion between the agricultural community and technology specialists is necessary
in order to recognize this symbiotic connection. Such collaboration acts as a link
between the concrete reality of on‑the‑ground farming practices and data‑driven
insights.
The effective integration of these two realms, where AI complements rather
than replaces human knowledge, is essential for contemporary agriculture. We can
achieve levels of efficiency, sustainability, and resilience in farming that have never
before been possible by creating a harmonic alliance between tradition and innova‑
tion. The preservation of ancestors’ wisdom and the use of cutting‑edge technology
are intertwined in this communal journey, offering a better future for agriculture.

11.5.3 Scaling AI Solutions for Smallholder Farmers

Smallholder farmers make up an important percentage of the global agricultural
population; they face resource constraints and limited access to cutting‑edge tech‑
nologies. Data‑driven AI must be both scalable and cost‑effective for it to benefit
everyone. As a result, affordable AI solutions must be developed that are especially
suited to the unique requirements and constraints of small‑scale agriculture.
The development of streamlined user interfaces that increase smallholder farm‑
ers’ access to AI tools is one strategy for reaching this objective. Additionally, as
mobile devices are frequently more common than conventional PCs in rural regions,
leveraging the power of mobile technology might be crucial in this endeavor. We can
provide farmers access to tools at their fingertips by creating AI programmers that
can operate on mobile devices.
Additionally, providing thorough training and assistance is essential. Smallholder
farmers might not be as comfortable with technology as larger businesses. Therefore,
focused training initiatives and continuing support can close this knowledge gap,
enabling these farmers to successfully embrace and apply AI solutions to improve
their farming methods. Democratizing AI in agriculture depends on making tech‑
nology accessible, inexpensive, and usable for smallholder farmers throughout the
world, who are the group that needs it most.
“Empowering smallholder farmers with AI solutions is not just about scaling
technology; it’s about scaling hope, resilience, and prosperity on the world’s most
important growth frontier.”
Role of Data‑Centric Artificial Intelligence in Agriculture 173

11.6 AI SUCCESS STORIES IN AGRICULTURE

Telangana is the success story of Indian agritech. AI tools, soil testing, e‑commerce &
more [34]: Telangana’s use of agritech is a shining example of an incredible achieve‑
ment in Indian agriculture. The state has used AI to transform the farming industry.
By 2025, the state hopes to offer agritech services to 100,000 farmers according to a
detailed plan it has established. This forward‑thinking strategy combines the estab‑
lishment of public infrastructure with smart partnerships, well‑designed initiatives,
and progressive legislation. The structure for public–private partnerships in the state
has encouraged collaboration between the federal and state governments and tech‑
nological firms, promoting the quick uptake of innovative solutions. Telangana has
also made the restructuring of the agro value chain, encouraging effectiveness and
sustainability in the industry.
Initiatives like the Agriculture Data Exchange (ADEx), which offers a platform
for experimentation and data exchange, and the Agritech Sandbox have fostered agri‑
tech innovation. Additionally, the creation of an agriculture data management system
guarantees that important data is tapped for well‑informed choices.
Success in agritech in Telangana goes beyond local achievement; it provides a
repeatable paradigm for other nations and regions seeking to digitally alter their agri‑
cultural industries. Telangana’s story provides a ray of light, illustrating technology
may be a driver for sustainable and wealthy agricultural futures as global agriculture
faces increasing difficulties.
AI‑Powered Agriculture: Revolutionizing Farming Practices [35]: Agerris, an
Australian start‑up, has transformed agriculture with its autonomous robot, Swagbot.
Using AI and ML, Swagbot navigates rugged terrain, analyzes crop data, and makes
real‑time decisions, increasing efficiency and profits for farmers. Agerris is also
pioneering AI systems for crop monitoring, disease detection, and yield prediction,
further enhancing farming practices. This success story showcases AI’s potential to
address global challenges like food security and climate change while optimizing
resource use. Agerris exemplifies how AI is revolutionizing agriculture, promising a
sustainable and prosperous future for farmers and the planet.

11.7 CONCLUSION
This chapter has explored the pivotal role of data‑centric AI in agriculture. The con‑
vergence of AI and agriculture is a promising frontier with enormous potential to
change how we think about modern farming. Effective use of AI technology depends
on an understanding of the significance of data in contemporary agriculture. We
explored the crucial components of data collection, highlighting the IoT’s contribu‑
tion to the collection of important data from agricultural activities. Once gathered,
this data goes through a substantial modification known as data preprocessing, which
makes it possible to use ML and DL algorithms designed particularly for agriculture
on it. DSSs have developed as an essential AI use in agriculture, offering farmers
insightful information and suggestions for improving their farming methods. We
also talked about the difficulties that must be overcome in order to fully realize AI’s
promise in agriculture, such as issues with data security and privacy, integration with
174 Data-Centric Artificial Intelligence for Multidisciplinary Applications

conventional agricultural expertise, and the requirement for scaling AI solutions for
smallholder farmers.
We also looked at real‑world examples of AI being successfully used in agri‑
culture, with Telangana serving as a shining example of how AI can improve agri‑
cultural practices. We also discussed the major advancements being made in this
industry by Australian start‑ups. The application of data‑centric AI in agriculture has
the potential to revolutionize the sector by presenting fresh approaches to venerable
problems. In order to ensure sustainable and effective agricultural production in the
years to come, it will be essential to handle the difficulties and seize the potential
given by AI in agriculture.

REFERENCES
1. Rozenstein, O., Cohen, Y., Alchanatis, V., et al. “Data‑driven agriculture and sus‑
tainable farming: friends or foes?” Precision Agric volume (2023). doi:10.1007/
s11119‑023‑10061‑5
2. Singh, P. “Systematic review of data‑centric approaches in artificial intelligence
and machine learning.” Data Science and Management 6, no. 3 (2023): 144–157.
doi:10.1016/j.dsm.2023.06.001.
3. Zha, D., Bhat, Z.P., Lai, K.H., Yang, F., Jiang, Z., Zhong, S., and Hu, X. “Data‑centric
artificial intelligence: a survey.” (2023). arXiv preprint arXiv:2303.10158.
4. Salehi, S., and Schmeink, A. “Data‑centric green artificial intelligence: a survey.” IEEE
Transactions on Artificial Intelligence. doi:10.1109/TAI.2023.3315272.
5. AlZubi, A.A., and Galyna, K. “Artificial intelligence and internet of things for
Sustainable Farming and Smart Agriculture.” IEEE Access 11 (2023): 78686–78692.
doi:10.1109/ACCESS.2023.3298215.
6. Elbasi, E. et al. “Artificial intelligence technology in the agricultural sector: a systematic
literature review.” IEEE Access 11 (2023): 171–202. doi:10.1109/ACCESS.2022.3232485.
7. Kobayashi, T., Yokogawa, T., Igawa, N., Sato, Y., Fujii, S., and Arimoto, K. “A Compact
low power AI module mounted on drone for plant monitor system.” In: 2019 8th
International Congress on Advanced Applied Informatics (IIAI‑AAI), Toyama, Japan,
2019, pp. 1081–1082. doi:10.1109/IIAI‑AAI.2019.00236.
8. Tantalaki, N., Souravlas, S., and Roumeliotis, M. “Data‑driven decision making in pre‑
cision agriculture: the rise of big data in agricultural systems.” Journal of Agricultural
& Food Information 20, no. 4 (2019): 344–380. doi:10.1080/10496505.2019.1638264
9. Wakchaure, M., Patle, B.K., and Mahindrakar, A.K. “Application of AI techniques and
robotics in agriculture: a review.” Artificial Intelligence in the Life Sciences 3 (2023):
100057. doi:10.1016/j.ailsci.2023.100057.
10. Dilmurat, K., Sagan, V., and Moose, S. “AI‑driven maize yield forecasting using
unmanned aerial vehicle‑based hyperspectral and lidar data fusion.” ISPRS Annals
of the Photogrammetry, Remote Sensing and Spatial Information Sciences 3 (2022):
193–199. doi:10.5194/isprs‑annals‑V‑3‑2022‑193‑2022.
11. U.S. Senate Committee on Agriculture, Nutrition, and Forestry. “Revisiting farm pro‑
duction expenses.” https://www.agriculture.senate.gov/newsroom/minority‑blog/revis‑
iting‑farm‑production‑expenses (accessed September 9, 2023).
12. ARK Invest. “Will the convergence between artificial intelligence and precision
agriculture lower farming costs?” ARK Invest (2023). https://ark‑invest.com/articles/
analyst‑research/will‑the‑convergence‑between‑artificial‑intelligence‑and‑preci‑
sion‑agriculture‑lower‑farming‑costs/ (accessed September 9, 2023).
Role of Data‑Centric Artificial Intelligence in Agriculture 175

13. Saiz‑Rubio, V., and Rovira‑Más, F. “From smart farming towards agriculture 5.0:
a review on crop data management.” Agronomy 10, no. 2 (2020): 207. https://doi.
org/10.3390/agronomy10020207.
14. Brown, A. “What is IoT in agriculture? Farmers aren’t quite sure despite $4bn US oppor‑
tunity‑report.” AgFunderNews (accessed September 2, 2023). https://agfundernews.
com/iot‑agriculture‑farmers‑arent‑quite‑sure‑despite‑4bn‑us‑opportunity.html.
15. Gralla, P. “Precision agriculture yields higher profits, lower risks.” HPE Insights
(June 2018). https://www.hpe.com/us/en/insights/articles/precision‑agriculture‑
yields‑higher‑profits‑lower‑risks‑1806.html (accessed September 2, 2023).
16. Tzounis, A., Katsoulas, N., Bartzanas, T., and Kittas, C. “Internet of things in agri‑
culture, recent advances and future challenges.” Biosystems Engineering 164 (2017):
31–48. doi:10.1016/j.biosystemseng.2017.09.007
17. Sarni, W., Mariani, J., and Kaji, J. “From dirt to data: the second green revolution
and IoT.” Deloitte Insights. https://www2.deloitte.com/insights/us/en/deloitte‑review/
issue‑18/second‑green‑revolution‑and‑internet‑of‑things.html (accessed September 2,
2023).
18. Mykleby, M., Doherty, P., and Makower, J. The New Grand Strategy: Restoring
America’s Prosperity, Security, and Sustainability in the 21st Century. St. Martin’s
Press (2016).
19. Shaikh, F.K., Karim, S., Zeadally, S., and Nebhen, J. “Recent trends in inter‑
net‑of‑things‑enabled sensor technologies for smart agriculture.” IEEE Internet of
Things Journal 9, no. 23 (2022): 23583–23598. doi:10.1109/JIOT.2022.3210154.
20. Shafi, U., Mumtaz, R., García‑Nieto, J., Hassan, S.A., Zaidi, S.A.R., and Iqbal, N.
“Precision agriculture techniques and practices: from considerations to applications.”
Sensors 19, no. 17 (2019): 3796. doi:10.3390/s19173796.
21. Nachankar, M. “Challenges of big data in agriculture: data collection and integra‑
tion.” International School of Advanced Management. https://isam.education/en/
challenges‑of‑big‑data‑in‑agriculture‑data‑collection‑and‑integration (accessed
September 2, 2023).
22. Wijaya, R., and Pudjoatmodjo, B. “An overview and implementation of extraction‑trans‑
formation‑loading (ETL) process in data warehouse (case study: department of agri‑
culture).” In: 2015 3rd International Conference on Information and Communication
Technology (ICoICT), 70–74. Nusa Dua, Bali, Indonesia, 2015. doi:10.1109/
ICoICT.2015.7231399.
23. Bocca, F.F., and Rodrigues, L.H.A. “The effect of tuning, feature engineering, and fea‑
ture selection in data mining applied to rainfed sugarcane yield modeling.” Computers
and Electronics in Agriculture 128 (2016): 67–76. doi:10.1016/j.compag.2016.08.015.
24. Sabarina, K., and Priya, N. “Lowering data dimensionality in big data for the benefit of
precision agriculture.” Procedia Computer Science 48 (2015): 548–554. doi:10.1016/j.
procs.2015.04.1345.
25. Yanwei, Y., Ling, X., Fuhua, J., Dafang, G., Sa, A., and Kang, N. “Experimental opti‑
mization of big data cleaning method for agricultural machinery.” Nongye Jixie Xuebao
Transactions of the Chinese Society of Agricultural Machinery 52, no. 6 (2021). https://
nyjxxb.net/index.php/journal/article/view/1190
26. Durai, S.K.S., and Shamili, M.D. “Smart farming using machine learning and deep
learning techniques.” Decision Analytics Journal 3 (2022): 100041. doi:10.1016/j.
dajour.2022.100041.
27. Saini, U., Kumar, R., Jain, V., and Krishnajith, M.U. “Univariant time series forecast‑
ing of agriculture load by using LSTM and GRU RNNs.” In: 2020 IEEE Students
Conference on Engineering & Systems (SCES), Prayagraj, India, pp. 1–6, 2020.
doi:10.1109/SCES50439.2020.9236695.
176 Data-Centric Artificial Intelligence for Multidisciplinary Applications

28. Lu, Y., Chen, D., Olaniyi, E., and Huang, Y. “Generative adversarial networks (GANs)
for image augmentation in agriculture: a systematic review.” Computers and Electronics
in Agriculture 200 (2022): 107208. doi:10.1016/j.compag.2022.107208.
29. Linaza, M.T. et al. 2021. “Data‑driven artificial intelligence applications for sustainable
precision agriculture.” Agronomy 11 (6): 1227. doi:10.3390/agronomy11061227.
30. Borrero, J.D., and Mariscal, J. “A case study of a digital data platform for the agricul‑
tural sector: a valuable decision support system for small farmers.” Agriculture 12, no.
6 (2022): 767. doi:10.3390/agriculture12060767.
31. Nikhil, R., Anisha, B.S., and Kumar, P.R. “Real‑time monitoring of agricultural land
with crop prediction and animal intrusion prevention using internet of things and
machine learning at edge.” In: 2020 IEEE International Conference on Electronics,
Computing and Communication Technologies (CONECCT), Bangalore, India, pp. 1–6,
2020. doi:10.1109/CONECCT50063.2020.9198508.
32. Lafont, M., Dupont, S., Cousin, P., Vallauri, A., and Dupont, C. “Back to the future:
IoT to improve aquaculture ‑ real‑time monitoring and algorithmic prediction of water
parameters for aquaculture needs.” In: 2019 Global IoT Summit (GIoTS), Aarhus,
Denmark, IEEE, 2019. doi:10.1109/GIOTS.2019.8766436.
33. Kumar, P., Gupta, G.P., and Tripathi, R. “PEFL: deep privacy‑encoding‑based feder‑
ated learning framework for smart agriculture.” IEEE Micro 42, no. 1 (2022): 33–40.
doi:10.1109/MM.2021.3112476.
34. Neo, G.H., and Rama Rao, K.T. “Telangana is the success story of Indian agri‑
tech: AI tools, soil testing, E‑commerce & more.” 2023. https://theprint.in/
economy/telangana‑is‑the‑success‑story‑of‑indian‑agritech‑ai‑tools‑soil‑testing‑
e‑commerce‑more/1630359/ (accessed September 2, 2023).
35. Ajayi, O. “AI‑powered agriculture: revolutionizing farming practices.” 2023. https://
www.linkedin.com/pulse/ai‑powered‑agriculture‑revolutionizing‑farming‑practices‑
ajayi/ (accessed September 2, 2023)
 12 Detection and
Classification of Mango
Fruit‑Based on Feature
Extraction Applying
Optimized Hybrid
LA‑FF Algorithms
Mukesh Kumar Tripathi, M. Neelakantappa,
Parikshit N. Mahalle, Shylesha V. Channapattana,
Ganesh Deshmukh, and Ghongade Prashant

12.1 INTRODUCTION
India is capable of producing a variety of horticulture products owing to its land
territory resilience. The overall horticultural output includes 90% of the fruit and
vegetables. The production of fruit and vegetables is 33% [1]. India is the leading
producer of mangoes. However, India is currently witnessing negative growth of
−0.86%. This is due to estimation loss during post‑harvest. Improper assessment,
wrong field h­ andling, transportation, mechanical damage during harvesting, and dis‑
ease cause quality losses in fruits. [2–5]. This is a serious issue that requires proper
attention. Mango fruits usually perish quickly, especially if stored in low tempera‑
tures of 7°C –13°C. Another cause for fruit losses is the traditional grading approach
[6,7]. This is more time‑consuming and labor intensive. This loss can be minimized
through proper framework and supply chain management with participants and other
entities. An optimal harvest time and the selection of quality are beneficial.
Customers power the fruit industry. The public trust in the fruit industry has been
diminished. Humans are more health conscious and vigilant. The definitions and
aspects of “quality” vary from the fruit class, the target audience, the requirements,
and the applicability. The assessment and grading of fruit quality is a progressively
complex task. In the past, determining fruit’s internal and external attributes was
challenging and time‑consuming. This is because of traditional evaluation methods
and a need for more research. The traditional approach uses tone to detect the quality

DOI: 10.1201/9781003461500-14 177

178 Data-Centric Artificial Intelligence for Multidisciplinary Applications

based on flavor. Yellow spots on the skin surface are another common form to iden‑
tify the disease. Much of the experiment also focused on external characteristics
and fruit deficiencies that could contribute to inaccuracies. Manual assessment relies
upon human activity. Research is based on the conventional technique for selecting a
high‑quality fruit in mango fruits [8]. In post‑harvest processing, this approach could
be more realistic. Demand for high‑quality fruit is growing, and non‑­destructive
automated techniques with a neutrosophic machine‑learning framework are also
desirable.
In today’s era, only a few studies have also been conducted for the quality evalu‑
ation of mangoes. Therefore, the suggested method is to grade the quality of mango
fruits based on external and internal characteristics. The destructive and non‑destruc‑
tive framework for quality grading is the most suitable solution. Our proposed system
can be further investigated for mango fruit quality assessment and grading with a
machine‑learning framework. Fruits are essential parts of the diet [9]. Fruits con‑
tain vital nutrients, fiber, energy, ascorbic acid, and proteins necessary for a healthy
human body [10]. Fruits are consumed in various forms as food or supplementary to
food. This mango is widely accepted due to its high nutrient value, taste, and flavor.
Mango is consumed in raw form or ripe form. The worldwide market for mango is
55 million tonnes [11].
High‑quality mangoes have increased day by day. Therefore, assessment and
grading are essential. Some internal features include soluble solids content (SSC),
total acid content (TAC), PH, physiological features, weight, dryness, firmness, mois‑
ture, and maturity [12–15]. Combining all these physical and biochemical param‑
eters defines the quality of mangoes. The grading based on external attributes could
be more efficient and accurate. Near‑infrared spectroscopy (NIRS) has excellent
potential for internal quality assessment and grading of mangoes. Further, to expand
the mango fruit market, it is necessary to develop an alternative framework to grade
the quality of mangoes [16–18]. This paper investigated the grading for mango fruit
quality with a neuromorphic approach‑based intelligent system. Then, hyperspectral
imaging is employed to estimate the internal attributes of the mangoes with machine
learning techniques.

12.2 RELATED WORK

Quality assessment is essential for a high‑quality product after harvest, as it is pos‑
sible to assess and grade fruit quality according to external and internal attributes.
Appearance is a critical perception in the assessment and grading of fruits. The
appearance will influence market value as well as consumer preference. However,
internal attributes are more essential to the grading of fruits. A methodology to assess
the quality of mango during the ripening process stage [19,20] is studied. The image
processing technique is applied for preprocessing and background subtraction. Also,
the ranked search method ranks the attributes of images belonging to information
gain. Finally, in the training and testing phases, C4.5 algorithms are implemented,
and a 96% accuracy rate is achieved overall. This framework is based on external
attributes. However, these studies have not considered internal attributes for grading.
Detection and Classification of Mango Fruit-Based on Feature Extraction 179

Maturity is one of the critical aspects of the quality of fruit. During the ripening
process, dry matter (DM) content is utilized to show the maturity level of mango
fruit. A robust and practical approach to recognize the spectral image and assess the
maturity level of mango fruit [21] is explored. They have implemented classification
and regression modeling to detect the mango and evaluate the quantity of DM. The
results show that for the Partial least squares regression (PLSR) model, R2 = 0.580 for
the CNN model. However, the system is unstable under natural light.
A random forest (RF)‑based model to evaluate the internal attributes of mango
[22] is investigated. The two categories of mango, namely “Nam Dakoi” and
“Irwin” at different temperatures are studied. L* a* b* color space is employed
to identify the color of the mango peel. The internal features such as Total soluble
solid (TSS) and ascorbic acid are calculated through destructive techniques. A RF
based model is utilized to grade the internal quality of mango. However, this model
has utilized destructive techniques.
NIRS techniques have gained attention as they are non‑destructive, fast, and
cost‑effective. NIRS‑based wavelength selection for calibration model [23] is studied.
This wavelength selection method has high consistency. Two different databases are
used to predict the effectiveness of the prediction model. The experiment combined
the wavelength selection method with standard sample calibration transfer methods.
The proposed method is applicable only in a high range of wavelengths of spectra.
A framework for the total acidity prediction of mango [24] by NIRS is presented.
This method utilized three regression approaches: Partial least squares regression
(PLSR), Support Vector Machine (SVM), and artificial neural networks (ANNs).
Spectra acquired in a wavelength range from 1000 to 2500 nm. Further, total acidity
is predicted. The calibration and prediction models achieved more than 90% accu‑
racy with the ANN model. However, handling spectra dimension is complex.
A hyperspectral imaging system to estimate moisture content [25] is studied.
Visible‑near infrared is applied for spectra with a wavelength range between 400
and 1000 nm, and second, NIR is employed with a wavelength range between 880
and 1720 nm. PLS is utilized for calibration and prediction models. Results show that
for mango samples in the spectral wavelength range between 400 and 1000 nm, the
accuracy rate is 43.70%, whereas an 87.15% accuracy rate has been achieved with
NIR‑based spectra wavelength range between 880 and 1720 nm.
Detection of firmness attribute framework for mango [25] based on NIRS is
explored. In this framework, both destructive and non‑destructive methodologies are
implemented to evaluate the internal quality of mangoes. Vis‑NIR extracts the spec‑
tra with a wavelength range of 400–1050 nm. Subsequently, the PLS model builds the
relationship between spectra and internal parameters. Further, genetic algorithms are
used to estimate the firmness level of mango fruit. Mango fruit image is a combina‑
tion of different internal biochemicals, and sometimes, it is difficult to estimate the
parameters based on a regression model. These challenges can be avoided by utiliz‑
ing NIRS with machine learning techniques.
The internal attributes of fruit are evaluated by applying destructive and
non‑destructive methods. Destructive approaches are time‑consuming and costly
in deciding the quality of the features, such as carotenoid materials, chlorophyll,
180 Data-Centric Artificial Intelligence for Multidisciplinary Applications

phenolic acids, and sugars. Research has therefore concentrated on developing and
applying non‑destructive techniques to quality assessment and evaluation of mature
stages in real time [26]. The benefit of non‑destructive techniques is that we can
track the same fruit for an extended period, refine calculations, and gain a more pro‑
found image of the actual properties of the fruit. These non‑destructive techniques
have been shown as efficient and proposed for evaluating quality grading in the fruit
industry [27].
NIRS has the advantage of extracting important internal attributes of fruits.
It has successfully tracked the production of disorders such as early detection
injury in the mango [28]. One limitation of the NIRS application is that it is
costly. However, NIRS is practical and usable for quality grading of fruits based
on internal features [29,30]. The critical problem of NIRS is the robustness of the
calibration mode [31] for fruit quality evaluation. Furthermore, the robustness
of NIRS models [32,33] often depends on fruit cultivation and harvest season.
The non‑destructive process with infrared spectroscopy is used to estimate the
internal attributes of the quality of the mango. The development of an automated
assessment and grading system is a complex task. In this framework, extracting
the feature of image data is difficult, followed by a training and classification

TABLE 12.1
Application of Machine Vision for Fruit Grading
Feature
Application Preprocessing Extraction Data Analysis Accuracy References
Sorting of Ostu Threshold Color and Fuzzy rule 94.97 [23]
mango Techniques Size
Grading of Binary Threshold Shape, Size, BPNN 80 [24]
date fruit Techniques Intensity
Grading of Binary Threshold Mass Statics 97 [25]
mango Techniques analysis
Sorting of Gamma curve fitting Color Coefficient of 98 [26]
mango determination
Sorting of Convolution Filter Color, ANN 80 [27]
mango volume
Grading of Threshold‑Techniques Size Caliber model 89.5 [28]
Mango
Grading of HSI Texture Neural 93.33 [29]
Mango Network
Grading of Binaries adaptive Color, PCA 92 [30]
Mango threshold shape
Classification Ostu Region Bayes 90.01 [31]
of mango Threshold‑Techniques classifier
Sorting of Threshold‑Techniques Mass and Regression 91.76 [32]
mango volume model
Detection and Classification of Mango Fruit-Based on Feature Extraction 181

process. With the above point, developing an accurate and efficient quality grad‑
ing model‑based approach is necessary.
After reviewing the data in Table 12.1, we observed that an accurate and efficient
quality grading model‑based approach is needed based on machine learning tech‑
niques. Developing innovation in consolidating the physicochemical and biochemi‑
cal data with machine vision conveys an effort towards a coordinated framework for
the agriculture industry. These goals will take care of the issues, yet they will like‑
wise give legitimate knowledge of internal and external parameters in the machine
vision framework.

12.3 PROPOSED MODEL

The diagrammatic progression of the proposed mango reviewing model is given in
Figure 12.1. This suggested work is chiefly expected to present a cleverly programmed
mango reviewing model with four stages, for example, (1) Preprocessing, (2) Element
Extraction, (3) Ideal Component Determination, and (4) Characterization. In the
pre‑handling stage, the accompanying system occurs, like picture resizing, commo‑
tion evacuation, and division. Picture resizing is essential to increment or lessen the
all‑out number of pixels. Commotion evacuation calculation is the method involved
with eliminating or diminishing the clamor from the picture. The picture is divided
into numerous sections, otherwise called picture objects.
The subsequent stage is component extraction, separating the shape, variety,
and surface highlights. Some of the conditions utilized in this work are minutes,
form region, shape edge, form guess, raised structure, looking at convexity, bounc‑
ing square shape most miniature encasing circle, fitting an oval, and fitting a line.

FIGURE 12.1 The diagrammatic flow of proposed mango grading model.

182 Data-Centric Artificial Intelligence for Multidisciplinary Applications

Before removing the condition, the info picture is entirely changed to a dim‑scale
picture. The variety highlights like histogram, mean, middle, standard deviation,
most excellent variety recurrence, and negligible variety recurrence are removed.
Before this component extraction, the RGB picture is changed over entirely to an
LAB (laboratory) image.

12.4 RESULTS AND DISCUSSION

The executed mango reviewing model is assessed utilizing Python. However, the
dataset comprises pictures of covered class names; we have coordinated to obtain 748
images of mangoes of various class marks. Among them, 169 views have a place with
solid mangoes, 34 sick mangoes, 192 ready mangoes, 164 unripe mangoes, 97 major
mangoes, 41 medium‑sized mangoes, and 49 highly enormous mangoes. The investi‑
gation is performed for the executed Streamlined Profound CNN model with ideal ele‑
ments over other ordinary models and regular highlights, and the outcomes are plotted.
Table 12.2 makes sense of the investigation of executed CNN against the regular
CNN model with all elements and the applicable highlights, Auto Encoder+ All
highlights RNN+ All highlights. The Unripe classification, the proposed model with
ideal highlights has achieved advancement for FDR (minimization measure), which
is 35.21% and 75.68% better than traditional CNN with all elements and regular
CNN with highlights in, Auto Encoder+ All endlessly includes RNN+ All elements
separately. The investigation subsequently made sense of the predominant exhibition
of the proposed CNN model with ideal elements against customary CNN.

TABLE 12.2
Comparison Performance Measure of All Categories of Mango with Respect
to All Method
Test CNN+ All CNN+ FIS Auto Encoder+ RNN+ All Optimized CNN+
cases Measures Features Features All Features Features Optimal Features
HD FOR 0.085 0.123 0.24 0 0.942
FPR 0.901 0 0 0.5117 0
FNR 0.2 1 0.8 0 0.9
FDR 0.871 0 0 0.587 0
RU FOR 0.087 0 0.24 0.57 0.0574
FPR 0.257 1 0 0 0.547
FNR 0.052 0 0.548 0.39 0.031
FDR 0.125 0.578 0 0 0.134
BMV FOR 0 0.1475 0.185 0.139 0.15
FPR 0.85 0 0 0 0
FNR 0 1 0.758 0.3433 0.7
FDR 0.758 0 0 0 0
Detection and Classification of Mango Fruit-Based on Feature Extraction 183

12.5 CONCLUSION
An automated grading system is designed to speed up the process of classifying the
mango images and facilitate quality evaluation in the industrial sector. A new hybrid
optimization algorithm, LA‑FF, is introduced to overcome the slow convergence.
The grading is evaluated based on healthy–diseased (HD), Ripe Unripe (RU), and
Big medium very big (BMV) categories, and in all test cases, our proposed method‑
ology achieves higher accuracy than conventional methods. Similarly, the proposed
Optimized CNN reaches the least False omission rate (FOR), False positive predic‑
tive value (FPR), False Negative predictive value (FNR), and False discovery rate
(FDR) values than the traditional method.

REFERENCES
[1] Nandi, C.S., Tudu, B., and Koley, C.: A machine vision technique for grading of har‑
vested mangoes based on maturity and quality, IEEE Sensors Journal 2016, 16(16),
6387–6396.
[2] Nambi, V.E., Thangavel, K., Jesudas, D.M.: Scientific classification of ripening period
and development of colour grade chart for Indian mangoes (Mangifera indica L.) using
multivariate cluster analysis, Scientia Horticulturae 2015, 193, 90–98.
[3] Tripathi, M.K., and Dhananjay D.M.: A role of computer vision in fruits and vegeta‑
bles among various horticulture products of agriculture fields: A survey, Information
Processing in Agriculture 2020, 7(2), 183–203.
[4] Anurekha, D., and Sankaran, R.A.: Efficient classification and grading of MANGOES
with GANFIS for improved performance, Multimedia Tools and Applications 2019, 79,
1–16.
[5] Wang, F., Zheng, J., Tian, X., Wang, J., and Feng, W.: An automatic sorting system for
fresh white button mushrooms based on image processing, Computers and Electronics
in Agriculture 2018, 151, 416–425.
[6] Tripathi, M.K., and Maktedar, D.D.: Recent machine learning based approaches for
disease detection and classification of agricultural products, In: 2016 International
Conference on Computing Communication Control and Automation (ICCUBEA),
IEEE, 2016, pp. 1–6.
[7] Bhatt, A.K., and Pant, D.: Automatic apple grading model development based on back
propagation neural network and machine vision, and its performance evaluation, AI &
Society 2015, 30(1), 45–56.
[8] Chiranjeevi, K., Tripathi, M.K., and Maktedar, D.D.: Block chain technology in agricul‑
ture product supply chain, In: 2021 International Conference on Artificial Intelligence
and Smart Systems (ICAIS), IEEE, 2021, pp. 1325–1329.
[9] Mohammadi, V., Kheiralipour, K., and Varnamkhasti, M.G.: Detecting maturity of per‑
simmon fruit based on image processing technique, Scientia Horticulturae 2015, 184,
123–128.
[10] Mohapatra, A., Shanmugasundaram, S., and Malmathanraj, R.: Grading of ripening
stages of red banana using dielectric properties changes and image processing approach,
Computers and Electronics in Agriculture 2017, 143, 100–110.
[11] Tripathi, M.K., and Dhananjay, D.M.: Optimized deep learning model for mango
grading: Hybridizing lion plus firefly algorithm, IET Image Processing 2021, 15(9),
1940–1956.
184 Data-Centric Artificial Intelligence for Multidisciplinary Applications

[12] Chen, X., Li, Z., Wang, Y., and Liu, J.: Effect of fruit and hand characteristics on
thumb‑index finger power‑grasp stability during manual fruit sorting, Computers and
Electronics in Agriculture 2019, 157, 479–487.
[13] Tripathi, M.K., and Dhananjay, D.M.: A framework with OTSUS thresholding method
for fruits and vegetables image segmentation, International Journal of Computer
Applications 2018, 975, 8887.
[14] Zhang, Y., Lee, W.S., Li, M., Zheng, L., Ritenour, M.A.: Non‑destructive recognition
and classification of citrus fruit blemishes based on ant colony optimized spectral infor‑
mation, Postharvest Biology and Technology 2018, 143, 119–128.
[15] Tripathi, M.K., and Maktedar, D.D.: Detection of various categories of fruits and veg‑
etables through various descriptors using machine learning techniques, International
Journal of Computational Intelligence Studies 2021, 10(1), 36–73.
[16] Zhang, Y., Wang, S., Ji, G., Phillips, P.: Fruit classification using computer vision and
feedforward neural network, Journal of Food Engineering 2014, 143, 167–177.
[17] Taghipour, A., and Frayret, J.‑M.: Coordination of operations planning in supply chains:
a review, International Journal of Business Performance and Supply Chain Modelling
2013, 5(3), 272–307.
[18] Channapattana, S.V., Srinidhi, C., Madhusudhan, A., Notla, S., Arkerimath, R., and
Tripathi, M.K.: Energy analysis of DI‑CI engine with nickel oxide nanoparticle added
Azadirachta indica biofuel at different static injection timing based on exergy. Energy
2023, 267, 126622.
[19] Alavi, N.: Quality determination of Mozafati dates using Mamdani fuzzy inference
system, Journal of the Saudi Society of Agricultural Sciences 2013, 12(2), 137–142.
[20] Gandomi, A.H., Yang, X.‑S., Talatahari, S and Alavi, A.H.: Firefly algorithm with
chaos, Commun Nonlinear Sci Numer Simulat 2013, 18, 89–98.
[21] Utai, K., Nagle, M., Hämmerle, S., Spreer, W., Mahayothee, B., Müller, J.: Mass estima‑
tion of mango fruits (Mangifera indica L., cv. Nam Dokmai) by linking image process‑
ing and artificial neural network, Engineering in Agriculture, Environment and Food
2019, 12(1), 103–110.
[22] Shivendra, K.C., and Tripathi, M.K.: Detection of fruits image applying decision tree
classifier techniques. In: Computational Intelligence and Data Analytics: Proceedings
of ICCIDA 2022, Springer Nature Singapore, Singapore, 2022, pp. 127–139.
[23] Saad, F.S.A., Ibrahim, M.F., Shakaff, A.Y.M., Zakaria, A., Abdullah, M.Z.: Shape
and weight grading of mangoes using visible imaging, Computers and Electronics in
Agriculture 2015, 115, 51–56.
[24] Schulze, K., Nagle, M., Spreer, W., Mahayothee, B., Müller, J.: Development and
assessment of different modeling approaches for size‑mass estimation of mango fruits
(Mangifera indica L., cv. Nam Dokmai), Computers and Electronics in Agriculture
2015, 114, 269–276.
[25] Tripathi, M.K., and Maktedar, D.D.: Internal quality assessment of mango fruit: an
automated grading system with ensemble classifier, The Imaging Science Journal 2022,
70(4), 253–272.
[26] Mizushima, A., Lu, R.: An image segmentation method for apple sorting and grad‑
ing using support vector machine and Otsus method, Computers and Electronics in
Agriculture 2013, 94, 29–37.
[27] Gurubelli, Y., Ramanathan, M., Ponnusamy, P.: Fractional fuzzy 2DLDA approach
for pomegranate fruit grade classification, Computers and Electronics in Agriculture,
2019, 162, 95–105.
[28] Nyalala, I., Okinda, C., Nyalala, L., Makange, N., Chao, Q., Chao, L., Yousaf, K., Chen,
K.: Tomato volume and mass estimation using computer vision and machine learning
algorithms: Cherry tomato model, Journal of Food Engineering 2019, 263, 288–298.
Detection and Classification of Mango Fruit-Based on Feature Extraction 185

[29] Tripathi, M.K., Maktedar, D., Vasundhara, D.N., Moorthy, C., and Patil, P.: Residual
life assessment (RLA) analysis of apple disease based on multimodal deep learning
model, International Journal of Intelligent Systems and Applications in Engineering
2023, 11(3), 1042–1050.
[30] Luo, F., Zhang, L., Du, B. and Zhang, L.: 2020. Dimensionality reduction with
enhanced hybrid‑graph discriminant learning for hyperspectral image classification.
IEEE Transactions on Geoscience and Remote Sensing, 58(8), 5336–5353.
[31] Tripathi, M.K., Neelakantapp, M., Nagesh Kaulage, A., Nabilal, K.V., Patil, S.N., and
Bamane, K.D.: Breast cancer image analysis and classification framework by apply‑
ing machine learning techniques, International Journal of Intelligent Systems and
Applications in Engineering 2023, 11(3), 930–941.
[32] LeCun, Y., Kavukvuoglu, K., and Farabet, C.: Convolutional networks and applications
in vision, In: Proceedings of 2010 IEEE international symposium on circuits and sys‑
tems, 2010, pp. 253–256.
[33] Boothalingam, R.: Optimization using lion algorithm: a biological inspiration from
lions social behavior, Evolutionary Intelligence 2018, 11(1‑2), 31–52.
Section III
Building AI with Quality Data
for Multidisciplinary Domains
 13 Solving Student
Guiding Your Way

Admission Woes
Snehal Rathi, Shekhar Chaugule,
Manisha Mali, Gitanjali R. Shinde,
and Swati Patil

13.1 INTRODUCTION
Our research paper focuses on addressing the challenges encountered by students
during the admission process for engineering colleges. The project, “Guiding Your
Way,” aims to alleviate the difficulties and provide a streamlined solution for stu‑
dents seeking admission to these institutions. We understand the complexities and
obstacles that students face during this critical phase and have developed this project
with the intention of making the process smoother and more efficient. The admis‑
sion process for engineering colleges entails enrolling on a dedicated website where
the percentage obtained in the diploma exams plays a crucial role. Once all students
have completed the enrollment, a comprehensive list is generated of the percentage
and rank of each student among all Maharashtra State Board of Technical Education
(MSBTE) candidates. The rank obtained becomes instrumental in determining the
cutoff list for various colleges. Our motivation for developing this project stemmed
from personal experiences, where we encountered similar challenges during our own
admissions. While I had a rank within the range of 700–750, allowing me to easily
identify the top ten colleges I could potentially be allotted to, some of my friends
faced a more daunting situation with ranks like 15,000 or 18,000. They were unsure
which colleges would consider their rank for admission. One significant hurdle they
faced was the cumbersome process of scrolling through a lengthy PDF document
containing hundreds of college listings. It became impractical and time‑consuming
to manually identify the colleges suitable for their rank. This inspired us to create a
solution that could provide a comprehensive list of colleges with a single click.
Our proposed solution involves the development of an application [1] that allows
students to input their preferred department, category, and admission criteria (rank or
percentage). By specifying a minimum and maximum rank, the application generates
a list of colleges with cutoffs falling within the given rank range. This feature not only
saves students from the anxiety and effort of manually searching through extensive
PDF documents but also significantly reduces the time required to obtain a suitable list
of colleges. Our project aims to alleviate the stress and uncertainty faced by students
during the admission process. By providing a user‑friendly platform that streamlines

DOI: 10.1201/9781003461500-16 189

190 Data-Centric Artificial Intelligence for Multidisciplinary Applications

the search for eligible colleges, we hope to empower students with the information
they need to make informed decisions about their educational future [2,3].

13.2 LITERATURE REVIEW

In this section, we describe the related concepts used in the specific literature and
also adapted in the proposed application.
CollegeDunia.com website reveals valuable insights into its features, services, and
impact on the college selection process. CollegeDunia.com is a popular platform
that provides comprehensive information about various colleges, courses, admission
processes, and other related details.
Additionally, the literature survey emphasizes the need for accurate and reliable
college predictions, especially for diploma students seeking admission to engineer‑
ing colleges. However, it is noted that CollegeDunia.com currently does not provide
college prediction services specifically for diploma students [4,5].
Careers360 is a prominent online platform that offers comprehensive informa‑
tion and guidance to students seeking admission to colleges in India. The website
provides extensive details about various colleges, courses, fees, admission processes,
and counseling services, making it a valuable resource for students and parents alike.
One notable feature of Careers360 is its extensive collection of student reviews and
ratings. These reviews provide firsthand experiences and opinions shared by cur‑
rent and former students, offering valuable insights into the quality of education,
facilities, and overall college environment. Career360.com is a valuable resource for
students seeking higher education in India. With its extensive database of colleges
and courses, the website provides students with reliable and accurate information to
make informed decisions regarding their academic pursuits. However, the college
prediction feature for diploma students is currently not available on the website and
will be provided on the Guiding Your Way website [6].
Shiksha.com is a renowned education portal in India that provides a comprehen‑
sive platform for students to explore and gather information about various colleges
and courses across multiple disciplines. The literature survey highlights the signifi‑
cance of Shiksha.com as a valuable resource for students seeking admission to medi‑
cal, engineering, design, and other fieldshiksha.com incorporates virtual assistants
into its system, enhancing the user experience and providing personalized guidance.
Virtual assistants [7,8], powered by artificial intelligence, offer real‑time assistance
to users, addressing their queries and helping them navigate the website effectively.
This feature demonstrates Shiksha.com’s commitment to leveraging technology to
support students in their educational pursuits [9,10].
One such website we analyzed was CollegeDekho.com, a popular platform that
offers comprehensive details about colleges, courses, and admission procedures.
However, during our analysis, we identified certain features that were not present in
CollegeDekho.com but were crucial for Diploma holder students. One prominent fea‑
ture lacking in CollegeDekho.com was the ability to find colleges based on rank and
caste, especially for Diploma holder students. While CollegeDekho.com provides
information about colleges and their cutoffs, it does not cater to the specific needs of
Diploma students who have unique criteria for admission. Diploma holders often face
Guiding Your Way: Solving Student Admission Woes 191

challenges in finding suitable colleges based on their rank and caste category, as their
admission process differs from that of other students.
To address this gap, our project, Guiding Your Way, focuses on incorporating this
essential feature. We have developed an application that allows Diploma holder stu‑
dents to easily search for colleges based on their rank range and caste category. This
feature streamlines the college selection process for Diploma students and provides
them with relevant and personalized options [11].
One key feature that is missing in IndiaCollegesHub.com is the ability to find
colleges based on rank and caste, especially for Diploma holder students. While the
website offers general information about colleges, it does not cater to the unique
requirements of Diploma students when it comes to college selection. Diploma hold‑
ers face challenges in finding suitable colleges based on their rank range and caste
category, which is different from the admission process for other students. Through
our literature survey, we have identified the limitation of IndiaCollegesHub.com in
meeting the specific needs of Diploma holder students. Our project aims to bridge
this gap and provide a comprehensive solution that allows Diploma students to find
suitable colleges based on their rank and caste category. By incorporating this fea‑
ture, Guiding Your Way offers a unique tool to facilitate the admission process for
Diploma holder students [12,13].
IndCareer.com primarily focuses on providing information about the hospital‑
ity of colleges, scholarships, and event details, along with general education sys‑
tem information [14]. However, our analysis revealed a significant gap in terms of
a specific feature that is vital for Diploma holder students. One crucial feature that
is absent in IndCareer.com is the ability to find colleges based on rank and castes
specifically for Diploma holder students. While the website offers valuable infor‑
mation about the hospitality sector and scholarships, it does not cater to the unique
needs of Diploma students when it comes to college selection. Diploma holders face
challenges in finding suitable colleges based on their rank range and caste category,
which is distinct from other students’ admission criteria [15].

13.3 TECHNOLOGY USED

The “Guiding Your Way” project utilized a range of technologies to develop an effi‑
cient and user‑friendly web‑based application for assisting students in the admission
process for engineering colleges. Python served as the primary programming lan‑
guage due to its versatility and extensive libraries. Flask, a lightweight web frame‑
work, was employed to handle the backend development, including routing and data
processing. To collect data from the result website, BeautifulSoup, a web scraping
library, was used in conjunction with Selenium, an automation tool for web brows‑
ers. This allowed for the extraction of relevant information from the website and its
subsequent conversion into an Excel spreadsheet format using the openpyxl library.
In this following technologies are used as follows:
Python scripts were then developed to further convert the data into JSON for‑
mat, enabling easier manipulation and integration with the application. HTML,
CSS, and JavaScript were used for the front‑end development, ensuring a visually
appealing and interactive user interface. The Git version control system was utilized
192 Data-Centric Artificial Intelligence for Multidisciplinary Applications

to manage the source code, facilitating collaboration and efficient tracking of code
changes. These technologies collectively enabled the development of a robust and
efficient web‑based application that streamlined the admission process for engineer‑
ing colleges.

• Python: Python programming language was used as the primary language

for the development of the “Guiding Your Way” project. Python’s versatil‑
ity, extensive libraries, and ease of use made it suitable for web scraping,
• Flask: Flask is a lightweight and flexible web framework in Python. It is
used to handle routing, request handling, and data processing.
• openpyxl: openpyxl is a Python library specifically designed for reading
and writing Excel files.
• HTML: Hypertext Markup Language (HTML) is the standard markup
language for creating web pages. It was used in conjunction with Flask to
develop the frontend interface.
• JavaScript: JavaScript is a widely used programming language that allows
for interactive and dynamic elements on web pages. It was utilized to add
interactivity and functionality to the “Guiding Your Way” application,
enabling features such as form validation and real‑time updates.
• JSON: JavaScript Object Notation (JSON) is a lightweight data interchange
format. After converting the collected data from the website into Excel format,
a Python script was developed to further convert it into JSON format. JSON
allowed for easier data manipulation and integration with the application.

13.4 METHODOLOGY
The college selection and admission process can be overwhelming for students due
to the vast amount of information available and the complexity of decision‑making.
To address these challenges, this research paper presents a proposed system called
“Guiding Your Way.” This innovative web‑based application aims to revolutionize the
college selection and admission process by providing students with accurate informa‑
tion, personalized recommendations, and efficient tools for decision‑making [8].
Comprehensive College Comparison: “Guiding Your Way” provides a compre‑
hensive college comparison feature that allows students to evaluate multiple colleges
based on various parameters such as infrastructure, faculty, placement records, and
academic programs. This enables students to make well‑informed decisions by con‑
sidering their specific preferences and priorities.

• Data Collection: The data required for the project was collected from the
result website, which provides information about students’ ranks and the
cutoff lists of various engineering colleges. The data collection process
involved scraping the website to extract the necessary information
• Data Conversion: The collected data was initially in a website format. To
analyze and manipulate the data effectively, it was converted into an Excel
spreadsheet format. Python programming language was utilized to develop
a web scraping script using libraries such as BeautifulSoup and Selenium to
extract the data from the website and store it in the Excel file (Figure 13.1).
Guiding Your Way: Solving Student Admission Woes 193

FIGURE 13.1 Converted dataset.

• Data Transformation: Once the data was in the Excel format, a Python
script using the openpyxl library was developed to convert the Excel
data into JSON format. This step was crucial as it allowed for easier data
manipulation and integration with the “Guiding Your Way” application
(Figure 13.2).

In the above system, work begins with data collection from the internet in the form
of PDF files. These files contain valuable information about colleges, including
rankings, courses, and other relevant details. To facilitate easy data manipulation,
a PDF‑to‑Excel converter is employed to convert the PDF files into an Excel for‑
mat. Once in Excel format, the system creates a dataset by organizing the data into
columns and rows. This dataset includes crucial information such as college names,
rankings, branches, and percentages. To enable further processing, an Excel to JSON
transformation is performed using a Python automation script. This script parses the
Excel file and converts the data into a JSON file, which offers a structured represen‑
tation suitable for filtering and querying.
When a user interacts with the system, they input their desired criteria, such as
rank, branch, and percentage. The system then applies filtering mechanisms to the
JSON data, narrowing down the options to colleges that meet the user’s specified
requirements.
194 Data-Centric Artificial Intelligence for Multidisciplinary Applications
Block diagram for the proposed system.
FIGURE 13.2
Guiding Your Way: Solving Student Admission Woes 195

After the system displays the college’s list.

• Application Development: The application development process involves

using a combination of programming languages and frameworks. Python,
Flask, and openpyxl were used to handle the backend functionalities of the
application. HTML, CSS, and JavaScript were employed for the front‑end
development to create an interactive and user‑friendly interface.
• Testing and Evaluation: To ensure the reliability and accuracy of the
“Guiding Your Way” application, extensive testing was conducted. This
included unit testing of individual components and integration testing to
verify the interaction between different modules. User testing was also per‑
formed to gather feedback and evaluate the usability of the application.
• Ethical Considerations: Throughout the project, ethical considerations
were taken into account, particularly regarding data privacy and security.
Measures were implemented to safeguard any personal or sensitive infor‑
mation collected during the data scraping process.
• Evaluation of Results: The results were evaluated by comparing the gen‑
erated JSON data with the original data from the website to ensure accu‑
racy and integrity. Additionally, the performance of the application was
assessed based on factors such as response time, user satisfaction, and sys‑
tem efficiency.

13.5 CHALLENGES IN THE PRESENT SYSTEM

The present system of education websites, such as Shiksha, CollegeDunia, and
Careers360, play a vital role in assisting students with information regarding col‑
leges, courses, and admission processes. However, despite their usefulness, these
platforms face several challenges that hinder their effectiveness. This research paper
aims to identify and address these challenges by proposing a novel system called
“Guiding Your Way.”
Limited Information: The existing education websites often provide limited infor‑
mation, making it difficult for students to make informed decisions regarding their
college choices. “Guiding Your Way” aims to overcome this challenge by offering
comprehensive and up‑to‑date information about colleges, including cutoff ranks,
courses offered, and eligibility criteria.
Inadequate Data Integration: The existing systems struggle with integrating data
from various sources, leading to inconsistencies and outdated information. The pro‑
posed system will address this challenge by employing robust data integration tech‑
niques to ensure accurate and real‑time data from reliable sources.
Inefficient College Comparison: The current websites often lack effective tools for
comparing different colleges based on various parameters such as infrastructure, fac‑
ulty, and placement records. The proposed system will incorporate a comprehensive
college comparison feature, enabling students to make informed decisions.
Inadequate Support for Diploma Students: Many existing platforms do not cater
specifically to diploma students seeking admission to engineering colleges. “Guiding
Your Way” will address this gap by providing specialized support and guidance for
diploma students, considering their unique requirements.
196 Data-Centric Artificial Intelligence for Multidisciplinary Applications

Lack of Career Guidance: The current education websites often overlook the cru‑
cial aspect of career guidance. The proposed system will integrate career guidance
resources, providing students with valuable insights into various career options asso‑
ciated with different courses and colleges.

13.6 RESULTS AND ANALYSIS

We present the results obtained from the evaluation of the Guiding Way system
and analyze its performance and effectiveness in addressing the challenges of the
admission process for engineering colleges. The analysis focuses on various aspects,
including data accuracy, system efficiency, and user satisfaction (Figure 13.3).
Study Content for the Syllabus of the Diploma Course:

• The Guiding Way project successfully retrieved study content for the syl‑
labus of the Diploma course from reliable sources.
Finding Best Colleges based on Percentage Range:
• The Guiding Way system effectively processed the percentage range pro‑
vided by students and generated a list of colleges matching the criteria.
• The algorithm compared the student’s percentage with the cutoff lists of various
colleges and identified the ones where admission was possible (Figure 13.4).

FIGURE 13.3 Home page.

Guiding Your Way: Solving Student Admission Woes 197

FIGURE 13.4 About us page.

We aimed to address the challenges faced by students during the admission process
for engineering colleges. We developed this project with the intention of solving
these problems and ensuring that no student encounters the same difficulties we
experienced. The admission process for engineering colleges involves enrolling
on a website and considering the aggregate percentage obtained in the final year
exams of the Diploma course. Once all students have enrolled, a comprehensive
list is generated, which includes the percentage and rank of each student among all
MSBTE students.
We observed that the use of rank is crucial in this process. By referring to the
previous year’s cutoff list, students can compare their rank and estimate the col‑
leges they may be allotted. However, we encountered a problem where students with
higher ranks, such as 15,000 or 18,000, struggled to identify the colleges they were
eligible for due to the large number of options listed in a PDF document.
To address this issue, we developed an application called Guiding Your Way.
This application allows students to select a particular department and their admis‑
sion category and choose between rank and percentage as the criteria. By entering a
minimum rank of 7000 and a maximum rank of 8000, for example, the application
generates a list of colleges with a cutoff between the specified rank range. This func‑
tionality saves students time and effort by providing a comprehensive list with just a
single click (Figures 13.5–13.7).

13.7 ADVANTAGE OF THE PROPOSED SYSTEM

OVER THE EXISTING SYSTEM
Firstly, Guiding Way provides a more streamlined and efficient process for students
seeking admission to engineering colleges. By automating the process of converting
data from the result website into a JSON file, Guiding Way eliminates the need for
198 Data-Centric Artificial Intelligence for Multidisciplinary Applications

FIGURE 13.5 Cutoff find page.

FIGURE 13.6 Result page.

students to manually sift through lengthy PDF documents. This not only saves time
but also reduces the chances of errors. Guiding Way offers personalized guidance to
students based on their rank and preferences. The system suggests a list of colleges
that the student is eligible for, based on their rank and desired course. This helps
students make informed decisions and increases their chances of getting admitted to
their preferred college. Guiding Way provides a user‑friendly interface that is easy to
navigate, even for those with limited technical knowledge. The system is designed
to be intuitive and user‑centric, with clear instructions and helpful tips. Guiding Way
Guiding Your Way: Solving Student Admission Woes 199

FIGURE 13.7 Contact us page.

is a comprehensive solution that addresses all the major challenges faced by students
during the admission process. From providing information on colleges to helping
students make choices based on their preferences, the system offers end‑to‑end guid‑
ance and support.

13.8 HARDWARE REQUIREMENTS

a. Computer: 256 GB HDD 4 GB RAM
b. Intel i3 processor

13.9 SOFTWARE REQUIREMENTS

a. Browser (Chrome, Mozilla, edge)
b. Vs.code

13.10 CONCLUSION AND FUTURE WORKS

In conclusion, the proposed system “Guiding Way” has successfully addressed the
challenges faced by students during the admission process for engineering colleges.
By leveraging web scraping techniques and data processing using Python, we have
developed a user‑friendly web application that simplifies the college selection process.
200 Data-Centric Artificial Intelligence for Multidisciplinary Applications

We have demonstrated the effectiveness of our system in providing accurate and

comprehensive information about college cutoff ranks and percentages. Our system
has proven to be a valuable tool for students, eliminating the need to manually search
through lengthy PDF documents and providing a streamlined interface to access rel‑
evant information. The inclusion of features such as filtering by rank or percentage
has further enhanced the usability of the application.
Overall, the “Guiding Way” system has demonstrated its potential to alleviate
the challenges faced by students during the admission process. Its accurate and effi‑
cient college selection mechanism, coupled with its user‑friendly interface, makes it
a valuable resource for students seeking admission to engineering colleges. Future
work could involve expanding the system’s database to include more colleges and
integrating real‑time updates for cutoff ranks and percentages to ensure up‑to‑date
information for users.

13.11 LIMITATION
However, we acknowledge several limitations in our current research. First, while
our project offers a useful application for students seeking admission to engineering
colleges, it does not introduce any novel methodologies. The lack of novel techniques
may limit the originality and potential impact of the work. Future research should
explore opportunities to incorporate innovative methodologies to further enhance the
project’s effectiveness and relevance. Secondly, “Guiding Your Way” currently relies
on data provided by the MSBTE enrollment website. Any discrepancies or inac‑
curacies in the data from the source may affect the application’s output. Therefore,
continuous efforts to validate and update the data sources are essential to ensure the
reliability of our application.

REFERENCES
[1] KularbPhettong, K. and Limphoemsuk, N. 2017. The effective of learning by augmented
reality on Android platform. In: E-Learning, E-Education, and Online Training,
Springer, Cham, pp. 111–118.
[2] Sahin, D. and Yilmaz, R.M. 2020. The effect of Augmented Reality Technology on mid‑
dle school students’ achievements and attitudes towards science education. Computers
& Education 144, 103710.
[3] Friendsickness in the Transition to College: Precollege Predictors and College
Adjustment Correlates. https://www.proquest.com/openview/8f3984886702e757f2d29
bcaa7276a9e/1?pq‑origsite=gscholar&cbl=18750
[4] This System Website Provide the Details about the College Site and Counseling System.
https://collegedunia.com/ and https://www.careers360.com/]
[5] Von Ah D, Ebert S, Ngamvitroj A, Park N, Kang DH. 2005. Predictors of health behav‑
iours in college students. Journal of Advanced Nursing 48(5),463-474. doi: 10.1111/
j.1365-2648.2004.03229.x. PMID: 15533084.
[6] Osadchyi, V.V., Valko, N.V. and Kuzmich, L.V., 2021, March. Using augmented reality
technologies for STEM education organizations. Journal of Physics: Conference Series
1840(1), 012027.
Guiding Your Way: Solving Student Admission Woes 201

[7] Fan, M., Antle, A.N. and Warren, J.L., 2020. Augmented reality for early language
learning: A systematic review of augmented reality application design, instructional
strategies, and evaluation outcomes. Journal of Educational Computing Research
58(6), 1059–1100.
[8] Ghulamani, S. and Zareen, S., 2018, March. Educating students in remote areas using
augmented reality. In: 2018 International Conference on Computing, Mathematics and
Engineering Technologies (iCoMET), IEEE, pp. 1–6.
[9] Rathi, S., Deshpande, Y., Nagaral, S., Narkhede, A., Sajwani, R. and Takalikar, V. 2021.
Analysis of user’s learning styles and academic emotions through web usage mining.
In: 2021 International Conference on Emerging Smart Computing and Informatics
(ESCI), Pune, India, pp. 159–164. doi:10.1109/ESCI50559.2021.9397037.
[10] Adrianto, D., Hidajat, M. and Yesmaya, V., 2016, December. Augmented reality using
Vuforia for marketing residence. In: 2016 1st International Conference on Game,
Game Art, and Gamification (ICGGAG), IEEE, pp. 1–5.
[11] Radosavljevic, S., Radosavljevic, V. and Grgurovic, B. 2020. The potential of imple‑
menting augmented reality into vocational higher education through mobile learning.
Interactive Learning Environments 28(4), 404–418.
[12] Aneena Aley Abraham (2023). Ranking System of Indian Universities and Colleges.
Retrieved February 12, 2024, from https://www.shiksha.com/science/ranking/
top-universities-colleges-in-india/121-2-0-0-0
[13] Gomes, C., Chanchal, S., Desai, T. and Jadhav, D., 2020. Class student management
system using facial recognition. ITM Web of Conferences 32, 02001.
[14] Raju, K.C., Yugandhar, K., Bharathi, D.V.N. and Vegesna, N., 2018, November.
Third based modern education system using augmented reality. In: 2018 IEEE 6th
International Conference on MOOCs, Innovation and Technology in Education
(MITE), IEEE, pp. 37–42.
[15] Živčić-Bećirević, I., Smojver-Ažić, S., & Dorčić, T. M. 2017. Predictors of university
students’ academic achievement: A prospective study. Društvena Istraživanja, 26(4),
457–476
 14 Melodic Pattern
Recognition for
Ornamentation Features
in Music Computing
Makarand Ramesh Velankar,
Sneha Kiran Thombre, and
Harshad Suryakant Wadkar

14.1 INTRODUCTION
Technology has made inroads in creative art domains such as music, drawing, etc.
Automatic pattern recognition in music is a challenging problem due to the complex
nature of music. Modeling and recognition of prosodic components such as ornamen‑
tation in music makes it more challenging. Examples of prosodic features used in
speech are tone, stressed word, or voice modulation. They are experienced in audio,
but cannot be represented in the corresponding text matter. Traditionally, different
pattern recognition (PR) paradigms for music pattern analysis include statistical and
structural approaches for a specific predefined task. A statistical approach based on
probabilistic models with efficient use of machine learning algorithms for different
applications in music information retrieval (MIR) is common among researchers.
The structural or symbolic approach based on formal grammar helps model melodic
or rhythmic structures in the music. The use of neural networks for PR now extends
to deep neural networks for efficient prediction. The boundaries between different
paradigms are fuzzy and fading. Combined approaches are also gaining popularity
as they share the same goals.
The statistical approach attempts to extract numerical values from the data as a
source for classification. This technique has gained more acceptance and popularity
in the research community due to different machine learning and classification algo‑
rithms applied to the numerical data extracted from the digital objects under study.
Self‑repetitive pattern identification is a topic of interest from music summary or
content‑based MIR. A comparative study of Chroma features, Constant Q transforms
features, and MFCC features was performed. Results were compared with ground
truth obtained from human expert annotation for identifying repetitive patterns [1].
The ground truth used in most of the systems is input from human experts.
It is not easy to get the ground truth for large datasets with duration in hours
for audio files. The challenge is to generate and evaluate ground truth for massive

202 DOI: 10.1201/9781003461500-17

Melodic Pattern Recognition for Ornamentation Features 203

musical data, which can evaluate different machine learning algorithms. The vector
space model was used for melodic pattern extraction of the raga in Indian art music,
and the results were assessed for diverse classification strategies [2]. The statistical
PR approach using vector representation is suitable for machine learning algorithms.
They typically require input data in feature values. Modeling appropriate features
of the music, which we interpret as structures, is the challenge in music computing.
The structural approach thus becomes necessary for melodic or rhythmic patterns,
perceived on a timeline as a sequence.
Structural or syntactic PR is generally applied for melodic or rhythmic PR. In
melodic patterns, the note sequence pattern is usually represented as an ordered list
of notes with string‑type data structure. A monophonic musical pattern is described
by using notes as the fundamental unit for the representation. The hierarchical tree
structure can be used to represent the pattern at different granularity levels. The
directed graph structure is another representation of the notes’ transition in melody.
Induction and matching of sequential melodic patterns pose several computational
challenges but are helpful for musically interesting retrieval tasks [3]. Time series
symbolic pattern representation of music is a challenge, as the data are multidimen‑
sional and real valued, with patterns rarely repeating precisely. Data margins and
fuzziness are used considering the perception of patterns. Pattern segmentation for
processing can be done using perceptually grouped melodic phrases. Pitch transcrip‑
tion, rhythmic meter, or tempo information may not be the only helpful information
for structural pattern analysis.
The use of timbre information with string‑matching techniques can be more ben‑
eficial for polyphonic music [4]. The music structural pattern representation plays a
significant role in the automatic conversion from a sheet printed music for the perfor‑
mance (interpretation) or vice versa (transcription). Evaluation of accuracies related
to optical music recognition systems used for structural musical interpretation is a
challenge [5]. A graphical structure can represent music scores. The terminal nodes
directly describe the content of the music, the internal nodes represent its incremen‑
tal summary, and the arc represents its relationship.
The similarity between two melodies can be calculated by analyzing the struc‑
ture of the graph and finding the shortest path between corresponding nodes in the
graph [6]. Due to the effective modeling of sequential patterns for the time‑variant
music and human perception of musical patterns, a structural approach is suitable for
melodic patterns. Therefore, more robust music knowledge representation systems
are possible with this paradigm. Human brains process musical patterns for different
interpretations, and a neural network‑based approach attempts to model the same.
The artificial neural network (NN) is a computing model composed of intercon‑
nections of artificial neuron units to simulate the human brain. Different problems
in music have been successfully addressed with the help of different variants of NN.
For example, the use of dynamic programming and recurrent NN with hidden states
or memory units for chord recognition is found better than the hidden Markov model
[7]. Convolutional Neural Network (CNN) trained on a mel‑scaled spectrogram
produced promising results for automatic boundary detection for musical structures
[8]. Research indicates that NN can be trained to identify statistical inconsistencies
across audio features to predict valence/arousal values for emotion classification [9].
204 Data-Centric Artificial Intelligence for Multidisciplinary Applications

FIGURE 14.1 Indian and Western music notation mapping.

Deep neural network models for solving PR problems are becoming increasingly
popular. They can primarily learn complex non‑linear input and output relationships,
have almost no dependence on domain‑specific knowledge, and the availability of
practical models. For the effective implementation of this strategy, a vast training
dataset is required. Datasets provide the foundation for machine learning Algorithms.
Indian music is represented using either Western or Indian notations. For example,
piano or harmonium as wind instruments with reeds representing the notes shows
similarity and mapping of both notations. The mapping is as shown in Figure 14.1.
The melodic patterns are described using a sequence of notes such as C D E in
Western form with equivalent Indian notations as Sa Re Ga. The notation representa‑
tion of songs is helpful for composers or performers to play songs during recording or
live shows. Traditionally, Indian classical music (ICM) uses various music ornamen‑
tation forms to convey specific rasa or mood during raga performance. Music orna‑
mentation is integrated into melodic and rhythmic patterns to add esthetic appeal
during the performance. Various ornamentation forms include kan swar, meend,
khatka, murkhi, Andolan, and gamaka. This ornamentation is used in Indian film
songs as a natural extension due to the influence of ICM. As per the inputs from
the performers [10], kan swar or grace note (tiny duration note) is used prominently
in Indian music for conveying emotional appeal. kan swar is sung or played before
or after the main note during the melody. A grace note is a similar notion used in
Western music ornamentation.
During the computational study of ornaments in Hindustani music [11], kan swar
or grace note was observed as a subtle change (delicate and difficult to analyze).
Other ornaments such as meend (glide from one note to another or glissando) or
krintan (kan swar followed by a meend) represent a variety of music ornaments used
in Indian music. The rate of change in frequency and amplitude constitutes ornamen‑
tation’s fundamental nature of the computational aspects.
The chapter is organized in the following manner. A detailed survey related to
data‑driven melodic PR and specific to music ornamentation feature is covered in
Section 2. Research gaps are presented in Section 3. Methodology along with an
exploratory learning algorithm used is explained in Section 4. Section 5 covers
results and discussions with future directions.
Melodic Pattern Recognition for Ornamentation Features 205

14.2 RELATED WORK

14.2.1 Data‑Driven Melodic PR
Data‑driven PR approaches are successfully applied in various domains. The detailed
survey has shown great promise for future developments of data‑driven models [12].
Data‑driven modeling using computational intelligence is likely to complement or
replace knowledge‑driven models [13]. Deep learning data‑driven models are success‑
fully implemented in domains such as healthcare [14], e‑commerce [15], bio‑medical
research for DNA sequence [16], and Cyber security [17]. In music computing, the
data‑driven approach is explored for various applications. Handwritten music notation
recognition [18], melodic PR [19], genre classification [20], music recommendation [21],
automatic transcription [22], and music generation [23] are some of the applications
explored by researchers using data‑driven approach. For melodic PR, attempts have
been made for tasks such as automatic chord recognition [24], F–0 or tonic estimation
[25], and vocal melody extraction [26]. For ICM, melody extraction using a data‑driven
approach is attempted for raga motif identification [27], raga recognition [28], emotion
classification [29], and genre classification [30]. For Indian music ornamentation fea‑
tures, the challenges of using the data‑driven model are presented by researchers [31].

14.2.2 Role of Music Ornamentation

Traditionally, Hindustani classical music uses various music ornamentation forms
to convey specific rasa or mood during raga performance. Various ornamentation
forms include kan swar, meend, khatka, murkhi, Andolan, and gamaka [32]. These
ornamentations are used in Indian film songs as a natural reflection because of the
classical music background of various composers and singers. As per the inputs from
the performers, kan swar (a tiny duration note) is used prominently in Indian music
for conveying emotional appeal.
Kan swar is sung or played before or after the main note during the melody.
During the Computational Study of Ornaments in Hindustani Music [11], it was
observed that, in kan swar presentation, the change is subtle (delicate and diffi‑
cult to analyze). Kan swar followed by a glide/meend from one note to another
(glissando) termed Krintan. Classification of melodic sequences into 17 different
clusters depending on various ornamentation features was proposed. It was also
proposed that the rate of change in frequency and amplitude domains together con‑
stitutes the fundamental nature of ornamentation. A grace note in Western music
is a similar notion as kan swar in Indian music. Grace note was considered an
Atemporal event in Western music with the addition of dummy silence if required
for alignment during real‑time music to score alignment [33]. During the study of
the timing of 11 grace notes in 45 piano performances, the grace notes with a longer
mean duration tended to deviate significantly from proportional duration, whilst
shorter grace notes were roughly invariant over tempo [34]. Features were numeri‑
cally extracted from score data in MusicXML and performance data in MIDI [35].
The number of notes preceding and following it is presented as a time signature 4D
vector as a feature for the grace note. The role of ornamentation features, especially
206 Data-Centric Artificial Intelligence for Multidisciplinary Applications

grace notes or kan swar, was prominent in conveying emotions [10]. Further, such
feature dimensions may be associated with different emotion classes. Therefore,
successfully capturing the expressive features of ornamentation and its association
with emotion can improve the accuracy of Music Emotion recognition classification.
Performers use ornamentation to provide a different experience to listeners during
the performances.

14.2.3 Music Ornamentation
Musical ornamentation is the way the performers or music arrangers add value to
the original musical composition to convey meaning [11]. Music ornamentation plays
a significant role in conveying the intended meaning and experience to the listener.
Different genres of music forms have varied ornamentation specific to the genre
with the cultural impact. Indian music consists of various musical ornaments such
as grace notes, vibrato, glides, and variations [32]. Many ornamentations use subtle
acoustic parameters such as timbre, intensity, and pitch. They are challenging to cap‑
ture, considering the nature of change and relativeness. In the case of Western music,
a majority of the ornaments are well‑documented with sheet music notations [33].
Western music performance guided by sheet music has a fixed format and relatively
less scope for improvisations. Jazz music has a similarity with Indian music consid‑
ering the improvisation aspect in both. In speech communication, pauses, stress on
different words, voice modulation, and repetition of specific words or phrases convey
content effectively. Similarly, in musical performance, these clues are used along
with musical embodiment, referred to as music prosody.
Music prosody is the non‑verbal clues used to convey musical meaning in the case
of vocal performances. In the instrumental version, the clues used are specific to the
instrument by utilizing the features and capabilities associated with the instrument.
Every performer attempts to convey the message in their style and interpretation.
One can notice that the performance by different artists for the same musical piece
or same artist at different times does have variations, and modeling these aspects is
a challenging task considering its subtle nature.
Indian music has various forms such as classical, regional, and traditional folk,
in popular. The ornamentation and prosodic elements used to differ with these
forms. Indian music is traditionally an orally transmitted teaching‑learning system,
and Pandit Bhatkhande and Pandit Paluskar introduced formal notations during
the beginning of the nineteenth century. It was probably the first sincere attempt of
music documentation in the form of notations which many musicians and composers
then adopted. These notations attempt to capture some of the ornamentation, such
as glides and vibrato which guides the performer. Thus, the performers have more
liberty within the musical framework of Indian music.

14.3 GAP ANALYSIS

As per the discussions with renowned performers [10], the grace note (kan svar) is
considered one of the critical ornamentations used across different music traditions
to convey emotions.
Melodic Pattern Recognition for Ornamentation Features 207

The gaps identified from the survey and discussions with researchers are as
follows:

1. As per the literature survey related to Indian music, it was observed that
very little work has been done so far to model and identify the music orna‑
mental features.
2. The annotated dataset for ornamental features is not readily available.

The need was felt to apply PR algorithms. For data‑driven approaches such as deep
learning, the annotated dataset required is quite large. Thus, the initial approach used
was exploratory learning with a small annotated dataset to identify the challenges
likely to be faced which need to be addressed for data‑driven methods.

14.4 METHODOLOGY
In our first attempt to explore the music embodiment of ornaments as prosodic fea‑
tures, we have developed an algorithm to capture this ornamentation feature using an
exploratory learning approach. The exploratory learning algorithm is developed to
identify ”grace note”, a source for inducing emotion as per performers. The experi‑
mentation was carried out with the help of annotations provided by domain experts.
Due to no fixed rules and boundaries documented for ornamentation with flexibility
for interpretations, the annotations provided by experts for the same musical piece
had some variations. A grace note is an ornamentation, defined as a short‑duration
note presented before or after a steady note. The duration of a grace note and steady
note is not fixed and is relative as per the perception or interpretations of domain
experts. Manual annotations are a time‑consuming process and have limitations due
to time availability by domain experts.
The annotated audio sample with waveform and annotated notes marked before
the main note is as shown in Figure 14.2. The audio sample has eight grace notes
sung before the main notes. The grace notes are of very small duration visible in

FIGURE 14.2 Annotated grace notes for audio sample.

208 Data-Centric Artificial Intelligence for Multidisciplinary Applications

the figure in blue line spikes. The main notes are represented in Indian, and their
counterpart Western note format is shown in brackets as SA (C4), Re (D4), respec‑
tively. That is, a note in Indian notation SA is shown in Western notation as C4
in the bracket. It can be noticed from the vertical bars in blue that the main note
durations are relatively long compared to a small duration grace note appearing
before them.
Grace notes (in bracket) annotated before the main notes in the Sargam as per
Indian notations are
(Re) Sa, (Ga)Re, (Ma)Ga, (Pa)Ma, (Dh)Pa, (Ni)Dh, (Sa’) Ni, (Re’) Sa’
The equivalent representation using western notations is
(D4) C4, (E4) D4, (F4) E4, (G4) F4, (A4) G4, (A#4) A4, (C5) B4, (D5) C5.
The following example demonstrates the grace notes identified using the devel‑
oped algorithm. Results are obtained with possible grace notes identified with the
note and duration. “gn” stands for the grace note identified, mentioned before the
grace note identified. Results show notes followed by relative durations and grace
notes as small relative durations, such as 1, 2, or 3. Whereas, the main notes have
longer relative durations, such as 13, 15, and 16.
[‘gn’, ‘ D4’, 2, ‘C4’, 13, ‘gn’, ‘E4’, 2, ‘D4’, 13, ‘gn’, ‘F4’, 3, ‘E4’, 16, ‘gn’, ‘G4’, 2,
‘F4’, 15, ‘gn’, ‘A4’, 2, ‘G4’, 15, ‘gn’, ‘A#4’, 2, ‘A4’, 16, ‘gn’, ‘C5’, 1, ‘B4’, 15, ‘gn’, ‘D5’,
2, ‘C5’, 16]
The result shows that eight grace notes are correctly identified out of a total of
eight grace notes annotated for the sample shown. The algorithm was further tested
for other samples. The results obtained are used to fine‑tune the algorithm to cor‑
rectly identify grace notes for different test samples. The results from 20 annotated
short samples were obtained using the developed algorithm and were discussed with
expert annotators for further improvements in the algorithm.
Algorithm for Grace note identification

// Initialization
I = 1//sample number initialized to the first sample
// Time instance of the sample I − T[I]
// Intensity of the sample I − A[I]
// Pitch of the sample I − P[I]
// Identification maximum Intensity and minimum intensity
MaxI = A[0]// Maximum Intensity audio sample initialization
i =1
while(End of samples i)// begin of loop 1
{
If(MaxI < A[i])
MaxI = A[i]
} // end of loop 1
MinI = MaxI− 30// Defines Audible Intensity range
// Intensity below MinI considered as silence
i =0
// Convert pitch info to notes N[ ] a ray
while(End of samples n)// begin of loop 2
{
Read P[i]
Melodic Pattern Recognition for Ornamentation Features 209

Identify the Western note(WN) from the pitch range//

Pitch range covers All audible frequencies
Write N[WN] // Notes in Western notation format which
covers an entire audible range
} // end of loop 2
i = 0\\ NSGN[] with grace notes identified
while(end of NewC)
{
Read 4 samples as NS[i], NS[i + 1], NS[i + 2] and NS[i + 3]
FN = NS[i]\\ first note
FD = NS[i + 1]\\ first note relative duration
SN = NS[i + 2]\\ second note
SD = NS[i + 3]\\ second note relative duration
if(FD < SD)
{
PD = 0.2 ∗ SD
}
if (FD < PD and FD < 3)\\ Grace note identification
{
Add NS[i], NS[i + 1], NS[i + 2] and NS[i + 3] to NSGN[]
Append NSGN[] GNP
i = i + 4\\ Previous note as grace notes
}
else IF(FD > SD and SD < 3)
{
ND = 0.2 ∗ FD
}
if(SD < ND)\\ Grace note
{
Add NS[i], NS[i + 1], NS[i + 2] and NS[i + 3] to NSGN[]
Append NSGN[] gn
i = i + 4\\ Next note as grace note
}
else
{
i = i + 2
}
}\\
NSGN[] contains grace note info

The samples used for training and testing the algorithm were containing grace notes
as the only ornamentation feature. Further, it was advised to test the algorithm for
audio samples with the presence of other ornamentation patterns such as glides. It
was observed that the algorithm provided the wrong output on a few occasions due to
overlap. Overlap of grace notes and confusion with glides was one of the significant
issues observed, as shown in Figure 14.3. Grace notes here are predicted as (A#4)
A4 and (C4) B4, which are observed as false positives compared with human annota‑
tions. Although the grace note is visible, it is referred to as glide by domain experts.
Thus, the algorithm misjudged the glide with grace note as it was not modeled for
other ornamentation features such as glide.
210 Data-Centric Artificial Intelligence for Multidisciplinary Applications

FIGURE 14.3 Grace notes confusion with glide.

The results showed reasonably good accuracy for samples; however, the algorithm
misinterpreted some instances. It is due to the overlap of ornamentation patterns such
as glides and grace notes.

14.5 RESULT AND DISCUSSION

Following were the critical observations from the result analysis and discussion with
the experts.

• Challenges observed to tackle false‑positive or true‑negative identification

cases.
• The role of intensity and timbral characteristics need to be studied.
• Pitch range considerations need to be revisited along with 12‑pitch octave
representation using the equal‑tempered scale.
• Need to study other ornaments such as glide, jerk, murki, Andolan vs.
grace note.
• Identification of grace notes in isolation without considering other orna‑
mentation leads to errors as glides are also identified as grace notes in some
cases.
• Data‑driven approach with unsupervised or semi‑supervised learning looks
promising approach to identify ornamentation patterns.

It was observed that although the algorithm developed has provided reasonably
good accuracy, it is still far away from practical use in applications. The rule‑based
approach has limitations for ornamentation patterns due to the complex nature of
the data. Experts did not annotate the grace notes identified by algorithms in some
cases. The algorithm failed to capture some grace notes, annotated by experts.
Non‑standardization of duration leads to some discrepancies among expert opinions.
Annotation and agreement among the experts is a challenging issue. The need felt to
standardize the duration of grace notes with experimentation and opinions of domain
experts. More experimentation and a huge annotated dataset are needed for enhanc‑
ing the algorithm and utilizing the ornamentation features for any application. Music
recommendation is a prominent application in music computing considering online
music consumption.
Melodic Pattern Recognition for Ornamentation Features 211

Annotating a large dataset for ornamental features is a very time‑consuming

task for humans. A data‑driven approach combining CNN and RNN may prove
to be a useful approach. CNN for the image analysis of the audio waveforms and
RNN with sequence modeling are suitable for ornamental feature identification.
Semi‑supervised or unsupervised or reinforcement learning approach to train a
model need to be explored with a data‑driven approach to identifying melodic pat‑
terns with respect to ornamentation features.
Deep belief network (DBN) is one of the best possible alternatives one can
explore. It has a stack‑restricted Boltzmann machine or RBM, which identifies hid‑
den patterns effectively in unsupervised learning methodology. The model can be
trained using a semi‑supervised learning approach with a small number of annotated
samples. The pretraining of the model makes it different from typical neural network
models. For music ornamentation features or patterns, DBN looks as a promising
alternative as having large labeled data for ornamentation patterns is very unlikely
considering the human efforts required. DBN with the feature of pretraining requires
small annotated data or labeled samples and thus it matches the requirement of the
suitable model for the task. The data‑centric or data‑driven model training happens
in the hidden layers of DBN and is the most promising approach identified for orna‑
mentation pattern identification in computational music. A data‑driven approach to
identify musical embodiment or ornamentation in the performance will be useful for
musicians and music learners.

REFERENCES
[1] Lu, L., Wang, M., and Zhang, H.‑J. (2004). Repeating pattern discovery and struc‑
ture analysis from acoustic music data. In: Proceedings of the 6th ACM SIGMM
international workshop on Multimedia information retrieval, New York, NY, USA,
pp. 275–282.
[2] Gulati, S., Serra, J., Ishwar, V., Senturk, S., and Serra, X. (2016). Phrase‑based raga
recognition using vector space modeling. In: 2016 IEEE International Conference on
Acoustics, Speech and Signal Processing (ICASSP), IEEE, pp. 66–70.
[3] Klapuri, A. (2010). Pattern induction and matching in music signals. In: Mitsuko
Aramaki, Mathieu Barthet, Richard Kronland‑Martinet, Sølvi Ystad (eds.)
International Symposium on Computer Music Modeling and Retrieval, Springer, New
York, pp. 188–204.
[4] Aucouturier, J.‑J., and Sandler, M. (2002). Finding repeating patterns in acoustic musi‑
cal signals: Applications for audio thumb nailing. In: Jyri Huopaniemi (ed.) Audio
Engineering Society Conference: 22nd International Conference: Virtual, Synthetic,
and Entertainment Audio, Audio Engineering Society.
[5] Bainbridge, D., and Bell, T. (2001). The challenge of optical music recognition.
Computers and the Humanities, 35(2), 95–121.
[6] Orio, N., and Roda, A. (2009). A measure of melodic similarity based on a graph repre‑
sentation of the music structure. In: ISMIR, Kobe, Japan, pp. 543–548.
[7] Boulanger‑Lewandowski, N., Bengio, Y., and Vincent, P. (2013). Audio chord recogni‑
tion with recurrent neural networks. In: ISMIR, Citeseer, pp. 335–340.
[8] Ullrich, K., Schl¨uter, J., and Grill, T. (2014). Boundary detection in music struc‑
ture analysis using convolutional neural networks. In: ISMIR, Taipei, Taiwan.
pp. 417–422.
212 Data-Centric Artificial Intelligence for Multidisciplinary Applications

[9] Vempala, N. N., and Russo. F. A. (2012). Predicting emotion from music audio fea‑
tures using neural networks. In: Proceedings of the 9th International Symposium on
Computer Music Modeling and Retrieval (CMMR), Lecture Notes in Computer Science,
London, UK, pp. 336–343.
[10] Personal Discussions (2012–2013). Personal discussions with renowned vocalist Veena
Saharabuddhe, Pandit Sanjeev Abhyankar and flute player Pandit Keshav Ginde.
[11] Narayan, A. A. (2018). Computational study of ornaments in Hindustani music. PhD
thesis, International Institute of Information Technology, Hyderabad.
[12] Bai, X., Wang, X., Liu, X., Liu, Q., Song, J., Sebe, N., and Kim, B. (2021). Explainable
deep learning for efficient and robust pattern recognition: A survey of recent develop‑
ments. Pattern Recognition, 120, 108102.
[13] Solomatine, D. P., and Ostfeld, A. (2008). Data‑driven modeling: some past experiences
and new approaches. Journal of Hydro Informatics, 10(1), 3–22.
[14] Chen, Y. W., and Jain, L. C. (2020). Deep Learning in Healthcare: Paradigms and
Applications, Springer, Heidelberg.
[15] Zhang, Q., Yang, L. T., Chen, Z., and Li, P. (2018). A survey on deep learning for big
data. Information Fusion, 42, 146–157.
[16] Busia, A., Dahl, G. E., Fannjiang, C., Alexander, D. H., Dorfman, E., Poplin, R., and
DePristo, M. (2018). A deep learning approach to pattern recognition for short DNA
sequences. BioRxiv, 353474.
[17] Sarker, I. H. (2021). Deep cybersecurity: a comprehensive overview from neural net‑
work and deep learning perspective. SN Computer Science, 2(3), 154.
[18] Calvo‑Zaragoza, J., Toselli, A. H., and Vidal, E. (2019). Handwritten music recogni‑
tion for mensural notation with convolutional recurrent neural networks. Pattern
Recognition Letters, 128, 115–121.
[19] Velankar, M., and Kulkarni, P. (2022). Melodic pattern recognition and similarity mod‑
eling: a systematic survey in music computing. Journal of Trends in Computer Science
and Smart Technology, 4(4), 272–290.
[20] Elbir, A., and Aydin, N. (2020). Music genre classification and music recommendation
by using deep learning. Electronics Letters, 56(12), 627–629.
[21] Van den Oord, A., Dieleman, S., and Schrauwen, B. (2013). Deep content‑based music
recommendation. In: C.J. Burges and L. Bottou and M. Welling and Z. Ghahramani and
K.Q. Weinberger (eds.) Advances in Neural Information Processing Systems, NeurIPS
Proceedings, 26.
[22] Benetos, E., Dixon, S., Duan, Z., and Ewert, S. (2018). Automatic music transcription:
An overview. IEEE Signal Processing Magazine, 36(1), 20–30.
[23] Briot, J. P., and Pachet, F. (2020). Deep learning for music generation: challenges and
directions. Neural Computing and Applications, 32(4), 981–993.
[24] Micchi, G., Kosta, K., Medeot, G., and Chanquion, P. (2021). A deep learning method
for enforcing coherence in automatic chord recognition. In: ISMIR, pp. 443–451.
[25] Bittner, R. M., McFee, B., Salamon, J., Li, P., and Bello, J. P. (2017, October). Deep
salience representations for F0 estimation in polyphonic music. In: ISMIR, Suzhou
(China) pp. 63–70.
[26] Su, L. (2018, April). Vocal melody extraction using patch‑based CNN. In: 2018 IEEE
International Conference on Acoustics, Speech and Signal Processing (ICASSP),
IEEE, Calgary, AB, Canada. pp. 371–375.
[27] Sharma, A. K., Aggarwal, G., Bhardwaj, S., Chakrabarti, P., Chakrabarti, T., Abawajy,
J. H., and Mahdin, H. (2021). Classification of Indian classical music with time‑series
matching deep learning approach. IEEE Access, 9, 102041–102052.
[28] Shah, D. P., Jagtap, N. M., Talekar, P. T., and Gawande, K. (2021). Raga recognition in
Indian classical music using deep learning. In: Artificial Intelligence in Music, Sound, Art
and Design: 10th International Conference, EvoMUSART 2021, Held as Part of EvoStar
2021, Proceedings 10, Virtual Event April 7–9, 2021, Springer, New York, pp. 248–263.
Melodic Pattern Recognition for Ornamentation Features 213

[29] Nag, S., Basu, M., Sanyal, S., Banerjee, A., and Ghosh, D. (2022). On the application of
deep learning and multi‑fractal techniques to classify emotions and instruments using
Indian Classical Music. Physica A: Statistical Mechanics and its Applications, 597,
127261.
[30] Singh, Y., and Biswas, A. (2022). Robustness of musical features on deep learning mod‑
els for music genre classification. Expert Systems with Applications, 199, 116879.
[31] Velankar, M., Thombre, S., and Wadkar, H. (2022). Evaluating deep learning models
for music emotion recognition. International Journal of Engineering Applied Sciences
and Technology, 7(6), 252–259.
[32] Pudaruth, S. K. (2016). A reflection on the aesthetics of Indian music, with special refer‑
ence to Hindustani raga‑Sangita. Sage Open, 6(4), 2158244016674512
[33] Cont, A. (2010). A coupled duration‑focused architecture for real‑time music‑to score
alignment. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(6),
974–987.
[34] Windsor, Luke, Aarts, Rinus, Desain, Peter, Heijink, Hank, and Timmers, Renee.
(2001). The timing of grace notes in skilled musical performance at different tempi: A
preliminary case study. Psychology of Music, 29(2), 149–169.
[35] Jeong, Dasaem, Kwon, Taegyun, Kim, Yoojin, and Nam, Juhan. (2019). Score and per‑
formance features for rendering expressive music performances. In: Music Encoding
Conference, Music Encoding Initiative Vienna, Austria, pp. 1–6.
 15 Content Analysis
Framework for Skill
Assessment
Abhishek Kabade, Harshad Jagadale,
Anurag Bharde, and S. P. Sonavane

15.1 INTRODUCTION
The education sector has undergone significant changes due to advancements in tech‑
nology, resulting in a need for innovative learning and skill assessment methods.
This book chapter delves into a cutting‑edge system that harnesses the power of the
Internet of Things (IoT), machine learning, artificial intelligence (AI), blockchain,
and RFID technology to address these challenges.
In today’s fast‑paced world, educational institutions must teach students theo‑
retical knowledge and practical skills that are relevant to the industry. Traditional
assessment methods may often fail to capture students’ abilities. Therefore, there is a
growing demand for a robust system that can accurately assess skills while fostering
collaboration between academia and industry. The system presented in this chapter
revolves around an AI‑based learning environment, incorporating an academic chat‑
bot and a central database to facilitate seamless communication between students
and industry professionals. The effectiveness of this system is demonstrated through
a case study conducted at an academic institution, Walchand College of Engineering
(WCE) in Sangli (MS), India.
At the core of the system lies the integration of RFID technology. Each student is
assigned a unique identity through an RFID card that stores comprehensive information,
including personal details, academic performance, and skill indices. These skill indices
represent the students’ abilities, encompassing both relative and absolute skill levels.
Stored in separate blocks within a blockchain network, these indices ensure data integ‑
rity and security while granting industry professionals access to relevant information.
The system also employs machine learning algorithms along with natural language pro‑
cessing techniques to extract meaningful content from diverse sources such as text and
audio materials. This content analysis enables the system to provide intelligent responses
to student queries and offer personalized recommendations for skill improvement.
The book chapter is divided into three parts to provide a comprehensive under‑
standing of the system.
Part I focuses on the infrastructure of the WCE campus network, highlighting
the academic chatbot, central database, and wireless RFID connections that form the
foundation of the system.

214 DOI: 10.1201/9781003461500-18

Content Analysis Framework for Skill Assessment 215

Part II delves into the intent behind industry queries, referencing query process‑
ing from Part I and introduces the content analyzer. The contents are selected from
documents, and for experimentation, the student resume document is taken as input
to extract skill‑indicating tokens resulting in Resume Analyzer [1]. This component
calculates skill indices based on the data stored on RFID cards. This section demon‑
strates how the system offers a comprehensive overview of students’ skills, assisting
both academia and industry in identifying suitable talent.
Part III explores the mechanism of recommending study materials and educa‑
tional resources tailored to each student’s specific skill improvement needs.
By integrating IoT, machine learning, AI, blockchain, and RFID technology, this
system offers a holistic approach to content learning and analysis for skill assess‑
ment. It promotes effective collaboration between academia and industry, enhances
skill evaluation, and provides personalized opportunities for skill development. The
subsequent sections of the chapter delve into the intricate details of the system, high‑
lighting its potential to revolutionize the educational landscape and meet the evolv‑
ing demands of the modern world.

i. Developed framework
ii. Connected with the cloud
iii. Integrated RFID
iv. AI/ML Block
v. Result Analysis

15.2 RELATED WORK

Chatbots have emerged as powerful conversational agents facilitating interactive and
personalized interactions between humans and computer systems. With their ability to
understand natural language input and generate contextually relevant responses, chat‑
bots have found applications in various domains, including customer service, virtual
assistants, and information retrieval systems. Chatbots [2] seem to hold great promise
when it comes to providing fast and convenient support to users by delivering targeted
answers to their questions. There’s a massive popularity for chatbots in the recent phase
within business organizations, technology firms, etc. which has been attributed to their
cost‑saving potential and has unlocked new abilities to handle multiple requests simulta‑
neously. Chatbots have completely evolved beyond being mere assistants and now have
established themselves as companions through their interactive responses and conversa‑
tion styles. According to a research study, customer service chatbots on various social
platforms often encounter requests that are emotional and sentimental to their health.
Many intelligent agents have emerged as powerful and highly promising tools for bet‑
ter mental health therapy [3]. Studies have promisingly demonstrated the effectiveness
of simple innovations such as engaging in deep breathing practices, acts of kindness,
increasing and improving positive emotions, and reducing negative emotions. Most of
the time [4], students have had to physically visit universities or colleges to collect vari‑
ous information like tuition fees, term schedules, and other relevant details during their
admission process or as per their daily needs. This process is very tedious and time‑­
consuming, also it requires manpower in providing the required information to visitors.
216 Data-Centric Artificial Intelligence for Multidisciplinary Applications

To address these challenges, a chatbot can be developed. The chatbot aims to facilitate
interaction between users and a chatbot that can be accessed anytime and anywhere.
By integrating the chatbot seamlessly into the university or college website through
simple language conversions, it becomes readily available to provide a wide range of
information related to the institution and student‑specific queries. The chatbot serves as
a valuable resource for anyone accessing the university’s website, allowing users to ask
questions pertaining to the university and receive corresponding responses generated by
an algorithm after processing the input message. Job search [5] from the online portal is
one of the most upcoming and efficient job search methods for both the job seeker and
the job provider. The solutions for these new‑age technologies are still the traditional
and time‑taking methods are still the same. The answers are driven by manual rules
like searching and reading the complete resume which takes a huge mental power and
time also hinging a bit with the effectiveness and frustration. Job finding is a type of
recommender system. The Recommender system was first introduced by Resnick and
Varian [6] who pointed out that in a typical recommender system, people provide rec‑
ommendations as inputs, which the system then aggregates and directs to appropriate
recipients. For job matching, many research works have been conducted to invent differ‑
ent recommender systems for job recruiting [7]. Among all of them, Malinowski et al.
[8] proposed bilateral matching recommendation systems for bringing people together
with jobs using an Expectation Maximization algorithm, while Golec and Kahya [9]
portray a fuzzy model for competency‑based employee evaluation and selection with
fuzzy rules. Paparrizos et al. [10] used Decision Table/Naive Bayes as a hybrid classifier.

15.3 PROPOSED SYSTEM

The overall system’s block diagram is demonstrated in Figure 15.1, where the
WCE network is connected to the cloud, facilitating the sharing and receiving of

FIGURE 15.1 Institutional level content learning and analysis for skill assessment.
Content Analysis Framework for Skill Assessment 217

information with industry professionals. Within the institutional campus network,

the server is connected to the database and also to the users, each of whom has their
own RFID. These RFID devices are linked to industry peers through the cloud, uti‑
lizing blockchain technology.

15.4 PART I: CHATBOT

This book chapter introduces a comprehensive methodology for harnessing natural
language processing (NLP) techniques in developing an academic chatbot as part of
an IoT‑based learning system.
NLP, a field within AI, aims to empower computers with the ability to compre‑
hend, interpret, and generate human language. By incorporating NLP techniques, the
chatbot becomes proficient in understanding the intent behind queries and deliver‑
ing meaningful responses. These techniques encompass various aspects of language
understanding and processing, including tokenization, stemming, named entity rec‑
ognition (NER) (identifying entities like names, organizations, and locations), senti‑
ment analysis (determining the expressed sentiment in text), and language modeling
(predicting word sequences).
By implementing NLP techniques in the academic chatbot, the IoT‑based learning
system offers students an interactive and intelligent interface for resolving queries
and receiving skill improvement recommendations. This integration enhances the
system’s ability to understand and process language, resulting in a more effective and
user‑friendly experience for students.

15.4.1 NLP Implementation
The implementation of the chatbot involves the use of Natural Language Toolkit
(NLTK), a widely adopted open‑source library in Python for NLP tasks. Figure 2
indicates the flow of implementation. The following steps outline the implementa‑
tion process:

1. Importing Corpus: This step entails accessing and uploading the data files
required to train and evaluate the NLP models. The system retrieves the
necessary data files, which serve as the dataset for training the chatbot. The
corpus may consist of various textual data such as documents, conversa‑
tions, or any relevant content.
2. Preprocessing Data: Data preprocessing is a critical step that involves
­cleaning and modifying the raw text to make it suitable for analysis. It
includes operations such as removing punctuation, converting all letters to
lowercase, eliminating stop words (commonly used words like “a,” “the,” and
“is” that carry little meaning), and handling special characters or symbols.
3. Test Case Processing: Test case processing involves preparing and manag‑
ing test cases to evaluate the chatbot’s functionality. Test cases comprise
input patterns and their corresponding expected responses, used to validate
the accuracy and validity of the chatbot’s replies. These test cases cover dif‑
ferent scenarios and user queries to ensure the chatbot performs as intended.
218 Data-Centric Artificial Intelligence for Multidisciplinary Applications

4. Tokenization: Tokenization involves breaking the text into smaller units

which are also known as tokens. Tokens typically represent words or
phrases, serving as the fundamental building blocks of the text. The tokeni‑
zation step segments the text based on specific rules or patterns, such as
whitespace or punctuation marks, to separate individual words or phrases.
Tokenization enables the chatbot to process and understand the meaning of
each token, facilitating further analysis.
5. Stemming: The technique of reducing words into their base or root forms
is called Stemming. It removes the suffixes or prefixes associated with the
words, focusing on the core part that represents the essence of the word.
Stemming helps consolidate words with similar meanings, reduces word
size, and improves the efficiency of text analysis tasks like sentiment analy‑
sis or information retrieval.
6. Bag of Words (BoW): The BoW model is a widely used representation
technique in NLP. It treats each document as a “bag” containing its words,
without considering grammar and word order. By constructing a vocabulary
of unique words in the corpus, the model represents each document as a vec‑
tor. Each dimension in the vector corresponds to a word in the vocabulary.
The values assigned to each dimension indicate the frequency or presence
of the respective word in the document. The BoW representation proves
valuable for various NLP tasks, including text classification, topic model‑
ing, and information retrieval.
7. One‑Hot Coding: One‑hot coding is a technique employed to numerically
represent categorical variables. In NLP, it can be used to represent words
or entities as binary vectors. Each word is assigned a unique index, and its
vector representation has a value of 1 for the corresponding index and 0
elsewhere.

By following these detailed steps, including data import, preprocessing, test case
processing, tokenization, stemming, BoW, and one‑hot coding, the NLP‑based chat‑
bot can effectively process and analyze textual data, generating precise and relevant
responses to user inquiries, depicted in Figure 15.2.

15.4.2 Block Diagram
15.4.3 Algorithm and Code: Chatbot Response Generation
Using TF‑IDF and Cosine Similarity
The code implements a chatbot that uses TF‑IDF and cosine similarity to find
the most similar sentence to the user’s response. It generates a response based on the
similarity, and if there is no understanding, it responds with a default message. The
TF‑IDF matrix is used to represent the importance of words in the sentences and to
measure the similarity between the user’s response and previous sentences cosine
similarity is used.
Content Analysis Framework for Skill Assessment 219

FIGURE 15.2 Block diagram for NLP implementation.

Algorithm

1. Start the response (user_response) function.

2. Initialize an empty string variable robo_response to store the chatbot’s
response.
3. Append the user’s response to the list of sentence tokens.
4. Create a TfidfVectorizer object with a custom tokenizer (lem_normalize)
and English stop words.
5. Compute the TF‑IDF matrix for all the sentence tokens using the vectorizer.
6. Calculate the cosine similarity between the user’s response and all other
sentences in the TF‑IDF matrix.
7. Find the index of the most similar sentence based on the cosine similarity.
8. Flatten and sort the cosine similarity values in ascending order.
9. Retrieve the second‑highest cosine similarity value.
10. If the similarity value is zero, set robo_response as “Sorry, I don’t under‑
stand you.”
11. Otherwise, retrieve the most similar sentence from the sentence_tokens list
using the obtained index and append it to robo_response.
12. Return the robo_response as the chatbot’s response.

Program Code

def response(user_response):
robo_response = ‘’
# Initialize the response variable for the chatbot
220 Data-Centric Artificial Intelligence for Multidisciplinary Applications

# Append the user’s response to the sentence tokens

sentence_tokens.append(user_response)
# Create a TF‑IDF vectorizer with a custom tokenizer and stop
words
vectorizer = TfidfVectorizer(tokenizer=lem_normalize,
stop_words=’english’)
# Compute the TF‑IDF matrix for the sentence tokens
tfidf = vectorizer.fit_transform(sentence_tokens)
# Compute the cosine similarity between the user’s response
and all other sentences
values = cosine_similarity(tfidf[‑1], tfidf)
# Find the index of the most similar sentence
idx = values.argsort()[0][‑2]
# Flatten and sort the cosine similarity values
flat = values.flatten()
flat.sort()
# Get the second highest cosine similarity value
req_tfidf = flat[‑2]
# Check if the similarity is zero, indicating no understanding
if req_tfidf == 0:
robo_response = ‘{} Sorry, I don\’t understand you’.
format(robo_response)
else:
# Retrieve the most similar sentence as the chatbot’s response
robo_response = robo_response + sentence_tokens[idx]
return robo_response

In the provided code snippet, cosine similarity is utilized to find the most similar
sentence to the user’s input. Cosine similarity is a metric which is used to measure
the similarity between two vectors, particularly in high‑dimensional spaces.
How it helps in generating responses to the chatbot:

1. Sentence Tokenization: The user’s input and other sentences are tokenized,
which means they are split into individual words or phrases.
2. TF‑IDF Vectorization: The TfidfVectorizer from the ‘sklearn.feature_
extraction.text’ module is used to convert the tokenized sentences to numer‑
ical representations called TF‑IDF vectors. TF‑IDF, which stands for Term
Frequency‑Inverse Document Frequency, is a technique commonly utilized
in information retrieval and text mining. Based on the frequency and rarity
of the words in the entire corpus, it assigns corresponding weights.
3. Cosine Similarity Calculation: The TF‑IDF vector for the user’s input
is compared with the TF‑IDF vectors of all the other sentences in the
­‘sentence_tokens’ list. The cosine similarity is calculated between the vec‑
tors using the ‘cosine_similarity’ function from the ‘sklearn.metrics.pair‑
wise’ module.
The cosine similarity between two vectors can be determined by evalu‑
ating the dot product of the vectors and dividing it by the product of their
Content Analysis Framework for Skill Assessment 221

FIGURE 15.3 Flowchart depicting the workflow of the chatbot.

magnitudes. Given two vectors A and B, the cosine similarity (similarity

score) is computed as follows:

Cosine Similarity = A·B / ( A * B ) (15.1)

where A · B represents the dot product of vectors A and B. ||A|| denotes

the magnitude (Euclidean norm) of vector A. ||B|| denotes the magnitude
(Euclidean norm) of vector B. The dot product of two vectors A and B is the
sum of the products of their corresponding components:
222 Data-Centric Artificial Intelligence for Multidisciplinary Applications

A·B = A1 * B1 + A2 * B2 + ... + An * Bn (15.2)

The magnitude (Euclidean norm) of a vector A is calculated as the square

root of the sum of the squares of its components:

|| A || = sqrt (A1∧2 + A2∧2 + ... + An∧ 2) (15.3)

By dividing the dot product of Vectors A and B by the product of their

magnitudes, the cosine similarity value is obtained. This value falls within
the range of –1 to 1, with 1 indicating strong similarity, 0 indicating no
similarity, and –1 indicating dissimilarity or oppositeness.
4. Finding the Most Similar Sentence: The index of the sentence with the
highest cosine similarity to the user’s input is obtained using the “argsort”
function. The response associated with that index is then selected as the
generated response for the chatbot.
By using cosine similarity, the chatbot identifies the sentence from the
existing dataset that is most similar to the user’s input. This allows the chat‑
bot to provide a response that is relevant and contextually appropriate to the
user’s query or statement

15.5 PART II: RESUME ANALYZER

The Resume Analyzer, which is discussed in Part 1 regarding query processing,
elevates skill assessment to a new level by leveraging the data stored on RFID cards
[11]. This section offers a detailed explanation of how the Resume Analyzer precisely
calculates skill indices based on the information extracted from the cards, thereby
providing a comprehensive overview of students’ skills.
The proposed system consists of two dashboards: one for users and another for
the admin. The system accepts resumes in PDF format, which are then extracted into
plain text. This extraction process utilizes the PDFMiner library to convert the PDF
resume into plain text format [12]. Next, Pyresparser is employed to extract keywords
from the resume, encompassing both technical and non‑technical terms. These key‑
words are stored in an array.
Additionally [13], the system allows the recruiter to specify their skillset require‑
ments, which are stored in a separate array. The system then matches these require‑
ments with the skills possessed by the candidates [14]. The percentage of requirements
fulfilled by each candidate is calculated by the system. Furthermore, the system
assigns a score to each uploaded resume based on various factors such as projects,
achievements, hobbies, positions of responsibility, and educational details mentioned
in the resume [15]. This score is intended to evaluate the overall quality and suit‑
ability of the resume.
In the context of resume analysis, the Resume Analyzer performs six essential
functions.
Content Analysis Framework for Skill Assessment 223

1. Standardization: A resume parser automatically extracts information

from resumes and transforms it into a consistent format, enabling effortless
searching, sorting, and comparison.
2. Keyword Extraction: This is a process performed by a resume parser that
identifies important keywords and phrases from a resume. These extracted
keywords and phrases can then be utilized to filter and rank candidates
according to their skills and qualifications.
3. Skills Matching: A resume parser matches the skills and qualifications
listed on the resume with the job description. This functionality assists
recruiters in identifying the candidates who are more suitable for the job
post based on the specific requirements [16].
4. Blockchain Integration: Blockchain technology can enhance the security
and immutability of the Resume Analyzer system. The candidate’s resume
data, including parsed information and Analyzed results, can be stored on
the blockchain network. This ensures that the information remains tamper‑
proof, transparent, and accessible only to the authorized parties.
5. Contact Extraction: A resume parser captures the candidate’s personal
details such as name, contact number, and email.
6. Qualification Matching: In the process of resume parsing, the system
extracts details related to the candidate’s work history, educational back‑
ground, certifications, and other pertinent qualifications, allowing for effec‑
tive qualification matching.
7. Data Analysis: Data obtained from resumes is analyzed to detect patterns
and trends, enabling organizations to enhance their recruitment procedures
and identify potential areas for optimization.

Algorithm of Pyresparser

Algorithm

13. Parse the resume document using the appropriate parser for the file format
(e.g., PyPDF2 for PDF).
14. Preprocess the parsed text to clean and normalize it.
15. Segment the preprocessed text into sentences.
16. Use spaCy’s NER module to identify entities (names, addresses, and phone
number) in each sentence.
17. Add the identified entities to the list of extracted entities.
18. Use pattern‑matching techniques to extract specific information (educa‑
tional qualifications, work experience, and skills) from the parsed text.
19. Add the extracted information to the list of extracted entities.
20. Convert the list of extracted entities into a structured format like JSON or
DataFrame.
21. Return the parsed resume information in the chosen structured format.
224 Data-Centric Artificial Intelligence for Multidisciplinary Applications

FIGURE 15.4 Flowchart illustrating the Resume Analyzer process.

Content Analysis Framework for Skill Assessment 225

15.5.1 Flow of Application

Application working shown in flowchart 15.4
The proposed model has five steps:

• Resume collections.
• Keywords searching and skills extraction.
• Matching skills with industry requirements.
• Calculating percentage requirement fulfilled by the candidates.
• Creating QR code of the uploaded resume and a self‑introductory video of
the user.
• Using Blockchain technology for data privacy and security

15.5.2 Resume Collection
Different machine learning algorithms are employed to suggest a shortlist of resumes
to recruiters from a large pool of resumes. However, these decisions rely on the
assumption that the data provided in the resumes is structured and standardized
within the same field.
For instance, if a company requires “Docker” skill, the system can search for the
keyword “Docker” in the resumes of the candidates. However, this method is not
able to give any insight into the candidate’s proficiency level in that specific field. It
is unable to determine the candidate’s level of expertise or competence in that par‑
ticular area.

15.5.3 Blockchain Implementation
In the proposed Figure 15.5, the process initiates with a request being initialized.
This request triggers the creation of a block that contains the student’s resume.
After the formation of the block, it is transmitted to peer industries within the
network. The block node then undergoes validation to ensure its authenticity and
integrity. If the block successfully passes the validation process, it is added to
the existing Blockchain, expanding the chain with the new block. Finally, the
request reaches the industry, marking the completion of the cycle. This diagram
visually represents the sequential flow of the request and highlights the crucial
steps involved in the creation and dissemination of blocks within the Blockchain
network.

15.6 PART III MECHANISM OF RECOMMENDING

EDUCATIONAL RESOURCES
This section explores the mechanism of recommending study materials and educa‑
tional resources tailored to each student’s specific skill improvement needs within
the IoT‑based learning framework.
226 Data-Centric Artificial Intelligence for Multidisciplinary Applications

FIGURE 15.5 Resume transaction through blockchain.

The algorithms are written in the Python language. These algorithms are used
for keywords searching and extraction. Extracted data is stored in MySQL database.
Stored data is fetched whenever required for matching with the given or mandatory
skill set of the industry. According to the candidate’s skills, the system will generate
his interested field. The framework suggests the candidate for the online courses and
YouTube channels as per his interest derived by the framework.
Based on his score, the framework classify them into three groups:

1. Fresher
2. Intermediate
3. Experienced

The score is calculated as the sum of percentage skills matched and the score
obtained for the mentioning other important details such as projects, achievements,
educational details, etc.
The candidates and companies will interact through web‑based interface devel‑
oped using “Streamlit” library. The percentage marks are divided by 2 to scale it in
terms of 50, and other 50 marks are for the mentioning of other important resume
details.
Content Analysis Framework for Skill Assessment 227

15.7 FLOWCHART OF QR CODE

FIGURE 15.6 Flowchart of QR code implementation using blockchain and RFID.

228 Data-Centric Artificial Intelligence for Multidisciplinary Applications

15.8 IMPLEMENTATION OVERVIEW

The work related to all the modules has been implemented using the web inter‑
face. Below are some of the modules. Streamlit library is used for web hosting
purposes. MySQL database is used to store data and fetch that data from the
database.
In Figure 15.7, Candidate resume. Here, he can upload his resume pdf.
Then, keywords are extracted from the resume and keywords are matched with
the required skill set of the company and the percentage requirement fulfilled is cal‑
culated as shown in Figure 15.8.
In Figure 15.9, the candidate can create his QR code for the link contain‑
ing resume soft copy with the self‑introductory video of Figure 15.10, the
Admin (Recruiter) get to know about all the uploaded resumes, their scores,

FIGURE 15.7 Collection of resumes and keywords extraction.

Content Analysis Framework for Skill Assessment 229

FIGURE 15.8 Percentage requirements fulfilled by the candidate.

recommended fields, name of the candidate, etc. Admin is also provided with a
pie chart showing recommended fields for the candidates depicting the percent‑
age share of each field.
In Figure 15.11, Company can add or delete skills from the skill set that the candi‑
date needs to have. Recruiter can add manually also. These skill set are then matched
with the skills of the candidate. The score is calculated from the percentage skills
matched with the company’s required skill set. The candidate’s extracted skill set and
the company’s skill set are stored in two different arrays. Their intersection is taken
to calculate the percentage score.
230 Data-Centric Artificial Intelligence for Multidisciplinary Applications

FIGURE 15.9 QR code generation.

FIGURE 15.10 Pie‑chart depicting the percentage of the candidate interested in each field.
Content Analysis Framework for Skill Assessment 231

FIGURE 15.11 Company’s required skill‑set figure enlarged.

15.9 REMARK
This book chapter explores an innovative IoT‑based learning system that integrates
machine learning, AI, blockchain, and RFID technology. The system aims to bridge
the gap between academia and industry by providing a comprehensive solution for
skill assessment and collaboration between the two sectors. The chapter discusses
the implementation details, with a particular focus on the development of an aca‑
demic chatbot by the use of NLP algorithms. The chatbot’s ability to get the language
and generate human‑like language enables meaningful interactions and personalized
recommendations for skill improvement.
The integration of blockchain technology in the Resume Analyzer module, along
with NLP techniques, enhances security, privacy, and trust in the resume verification
process. It enables efficient text processing and addresses the challenges associated
with natural language extraction. However, the system may face difficulties in disam‑
biguating similar words with different meanings. Despite this, it serves as a primary
tool for identifying eligible candidates for job opportunities.
232 Data-Centric Artificial Intelligence for Multidisciplinary Applications

Additionally, the use of QR codes for recruiters streamlines the hiring process,
saving time and providing convenience. However, it is important to carefully evaluate
the challenges and limitations associated with blockchain technology in the resume
analysis system. Factors such as complexity, scalability, and integration efforts
should be considered based on the specific requirements and goals of the application.

15.10 FUTURE WORK

The future work in the Chatbot and Resume Analyzer system aims to focus on sev‑
eral key factors, including scalability enhancement, network capability improve‑
ment, block integrity, training set enhancement, and evaluation of proficiency of the
candidate in a particular field.

1. Scalability Enhancement: Using various methods such as shading, side‑

chains, or Layer‑2 solutions to enhance the system’s capacity in handling an
increasing number of resumes and users.
2. Network Capability Improvement: Collaboration with network engi‑
neers is undertaken to upgrade hardware, improve network protocols, and
ensure the seamless and efficient transmission of data between network
components.
3. Block Integrity Assurance: Using hashing algorithms in conjunction with
digital signatures allows us to verify the authenticity of resumes and prevent
any tampering or unauthorized modifications.
4. Proficiency Assessment: If a candidate mentions “Java” in their resume,
our system can determine that they have knowledge of “Java”. However,
the system does not estimate their proficiency level. To address this, we can
develop a chatbot function integrated with the Resume Analyzer. This chat‑
bot function will interact with the candidate and evaluate their proficiency
in specific technical keywords.

REFERENCES
1. Resume Screening Using LSTM, International Journal of Research Publication and
Reviews, 3(4), 2567–2569, https://ijrpr.com/uploads/V3ISSUE4/IJRPR3705.pdf
2. Adamopoulou, E., and Moussiades, L. (2020). An overview of chatbot technol‑
ogy. In: Maglogiannis I., Iliadis L., Pimenidis E. (eds), Artificial Intelligence
Applications and Innovations. AIAI 2020. IFIP Advances in Information and
Communication Technology, vol. 584, Springer, Cham. https://www.researchgate.net/
publication/341730184_An_Overview_of_Chatbot_Technology
3. Ghandeharioun, A., McDuff, D., Czerwinski, M., and Rowan, K. (2019). EMMA:
An emotion‑aware wellbeing chatbot. In: 2019 8th International Conference on
Affective Computing and Intelligent Interaction (ACII), Cambridge, UK, 2019, pp. 1–7,
doi:10.1109/ACII.2019.8925455. https://ieeexplore.ieee.org/abstract/document/8925455
4. Patel, N. P., Parikh, D. R., Patel, D. A., and Patel, R. R. (2019). AI and web‑based
human‑like interactive university chatbot (UNIBOT). In: 2019 3rd International
Conference on Electronics, Communication and Aerospace Technology (ICECA),
Coimbatore, India, 2019, pp. 148–150, doi:10.1109/ICECA.2019.8822176. https://ieeex‑
plore.ieee.org/abstract/document/8822176
Content Analysis Framework for Skill Assessment 233

5. Lin, Y., Lei, H., Addo, P. C., and Li, X. (2016). Machine learned resume‑job matching
solution. https://arxiv.org/abs/1607.07657
6. Resnick, P., and Varian, H. R. (1997). Recommender systems. Communications of the
ACM, 40(3), 56–58. doi:10.1145/245108.245121
7. Al‑Otaibi, S. T., and Ykhlef, M. (2012). A survey of job recommender systems.
International Journal of the Physical Sciences, 7(29), 5127–5142. https://academicjour‑
nals.org/journal/IJPS/article‑full‑text‑pdf/B19DCA416592.pdf
8. Malinowski, J., Keim, T., Wendt, O., and Weitzel, T. (2006). Matching people and
jobs: A bilateral recommendation approach. https://www.researchgate.net/publi‑
cation/232615527_Matching_People_and_Jobs_A_Bilateral_Recommendation_
Approach
9. Golec, A., and Kahya, E. (2007). A fuzzy model for competency‑based employee evalu‑
ation and selection. Computers & Industrial Engineering, 52(1), 143–161. https://www.
researchgate.net/publication/222332549_A_fuzzy_model_for_competencybased_
employee_evaluation_and_selection
10. Paparrizos, I., Cambazoglu, B. B., and Gionis, A. (2011). Machine learned job rec‑
ommendation. In: Proceedings of the Fifth ACM Conference on Recommender
Systems, ACM, New York, pp. 325–328. https://www.researchgate.net/publication/
221141098_Machine_learned_job_recommendation
11. Bojars, U., and Breslin, J. G. ResumeRDF: Expressing skill information on the semantic
web. https://www.researchgate.net/publication/266448089_ResumeRDF_Expressing_
skill_information_on_the_Semantic_Web
12. (2015). Resume analyzer an automated solution to recruitment process. International
Journal of Engineering nd Technical Research, 3(8). https://www.erpublication.org/
published_paper/IJETR032886.pdf
13. Bojars, U. Extending FOAF with resume information. https://www.w3.org/2001/sw/
Europe/events/foaf‑galway/papers/pp/extending_foaf_with_resume/
14. Verma, M. (2017). Cluster based ranking index for enhancing recruitment process using
text mining and machine learning. International Journal of Computer Applications,
157(9). https://www.researchgate.net/publication/312518297_Cluster_based_Ranking_
Index_for_Enhancing_Recruitment_Process_using_Text_Mining_and_Machine_
Learning
15. Jurka, T. P., Collingwood, L., Boydstun, A. E., Grossman, E., and van Atteveldt, W.
(2013). A supervised learning package for text classification. The R Journal. https://
journal.r‑project.org/archive/2013/RJ‑2013‑001/index.html
16. Li, L., Chu, W., Langford, J., and Schapire, R. E. (2010). A contextual‑bandit approach
to personalized news article recommendation. In: Proceedings of the Nineteenth
International Conference on World Wide Web. https://arxiv.org/pdf/1003.0146.pdf
 16 Machine‑Learning
Techniques for
Effective Text Mining
Shivam Singh, Chandrakant D. Kokane,
Vilas Deotare, and Tushar Waykole

16.1 INTRODUCTION TO TEXT MINING

AND MACHINE LEARNING
16.1.1 Definition and Scope of Text Mining
Text mining is the process of collecting valuable information, patterns, and insights
from unstructured textual data. It is also known as text data mining or text analyt‑
ics. Unstructured text data includes a wide range of sources such as emails, social
media posts, news articles, research papers, customer reviews, and more.1 Unlike
structured data, which is organized in a predefined format (e.g., databases, spread‑
sheets), unstructured text lacks a clear and uniform structure, making it challenging
to analyze using traditional methods.
The Definition of Text Mining: Text mining involves the application of vari‑
ous computational techniques, natural language processing (NLP) algorithms, and
machine‑learning methods to process, understand, and analyze unstructured textual
data. The ultimate goal of text mining is to convert the raw text into structured,
actionable information, enabling users to uncover hidden patterns, trends, senti‑
ments, relationships, and other valuable insights within the data.2
The Scope of Text Mining: The scope of text mining is extensive, and it finds
applications in diverse fields due to the widespread availability of textual data. Some
key areas where text mining is applied include:

1. Information Retrieval: Text mining is used to build search engines and

information retrieval systems that can efficiently retrieve relevant docu‑
ments or information based on user queries.
2. Sentiment Analysis: Sentiment analysis, often known as opinion mining,
is a subset of text mining that focuses on determining a text’s sentiment or
emotional tone. It is extensively used in social media monitoring, market
research, and client feedback analysis.
3. Text Classification: Text mining allows for the classification of texts based
on their content into predetermined groupings or categories. This has uses
in spam identification, topic tagging, document organizing, and other areas.

234 DOI: 10.1201/9781003461500-19

Machine‑Learning Techniques for Effective Text Mining 235

4. Named Entity Recognition (NER): NER is the process of identifying

and categorizing named entities in text, such as people’s names, organiza‑
tions’ names, locations, dates, and other specified phrases. It is useful in the
extraction of information and the building of knowledge graphs.
5. Topic Modeling: Topic modeling is a technique for identifying and extract‑
ing hidden topics or themes in a big collection of texts. It finds use in content
recommendation systems, content analysis, and document clustering.
6. Language Translation: Text mining techniques can be used in machine
translation systems to automatically translate text from one language to
another.
7. Text summarizing: Automatic text summarizing is the process of produc‑
ing succinct summaries of lengthy documents or articles, allowing users to
grasp the important points without having to read the entire text.
8. Fraud Detection: Text mining can be applied in fraud detection and pre‑
vention, such as identifying suspicious text patterns in emails or financial
reports.
9. Healthcare and Biomedical Applications: Text mining plays a vital role
in extracting useful information from medical records, research articles,
and clinical notes, enabling better decision‑making in healthcare.
10. News Analysis: Text mining is used to scan news articles and social media
feeds in order to find patterns, sentiment around specific topics, and public
opinion.

The scope of text mining is continually expanding as researchers and practitioners

develop new techniques and adapt existing ones to address emerging challenges and
domains. However, it’s important to note that text mining also presents certain chal‑
lenges, such as dealing with noisy data, language complexities, privacy concerns, and
ethical considerations, which need to be carefully addressed in its applications.4 The
operational process of text mining is illustrated in Figure 16.1.

16.1.2 Importance of Machine Learning in Text Mining

Machine learning is critical in text mining because it allows for effective analysis and
extraction of useful insights from unstructured text data.1 The following major char‑
acteristics help to understand the significance of machine learning in text mining:

1. Handling Unstructured Data: Text data is typically unstructured and

lacks a predefined format, making it challenging to process and analyze
using traditional methods. Machine‑learning algorithms are well‑suited to
handle unstructured data and can automatically learn patterns and struc‑
tures from textual information, transforming it into a structured format for
further analysis.
2. Feature Extraction and Representation: Machine‑learning techniques
provide various feature extraction and representation methods that help
convert raw text into meaningful numerical representations.2 These meth‑
ods, such as Bag‑of‑Words (BoW), Term Frequency‑Inverse Document
236 Data-Centric Artificial Intelligence for Multidisciplinary Applications

FIGURE 16.1 Text mining working.

Machine‑Learning Techniques for Effective Text Mining 237

Frequency (TF‑IDF), and Word Embeddings, capture the semantic mean‑

ing and context of words, making it easier for algorithms to understand and
analyze textual data effectively.
3. Text Classification and Categorization: Machine‑learning algorithms
excel in text classification tasks where documents are categorized into
predefined classes or categories. Whether it’s categorizing news articles,
sentiment analysis of customer reviews, or spam detection in emails,
machine‑learning models can automatically learn from labeled data and
classify new text data accurately.
4. Sentiment Analysis and Opinion Mining: Extracting sentiment and
views from text is a difficult undertaking that necessitates a thorough grasp
of the context and emotional tone. SVMs and Recurrent Neural Networks
(RNNs) are strong sentiment analysis models that allow businesses to
glean useful insights from consumer feedback, social media posts, and
other sources.
5. Named Entity Recognition (NER): In NER, machine‑learning approaches
are frequently used to detect and categorize named entities in text, such as
names of people, organizations, locations, and dates. NER is required for
a variety of applications such as information extraction, knowledge graph
generation, and data connecting.
6. Topic Modeling: Topic modeling methods, such as Latent Dirichlet
Allocation (LDA) and Non‑negative Matrix Factorization (NMF), are a
subset of machine learning that aid in the discovery of hidden topics or
themes within a huge collection of texts. This aids in the organization and
comprehension of big text corpora, and it is useful for content suggestion
and analysis.
7. Text Clustering: Unsupervised machine‑learning methods such as
K‑means and Hierarchical Clustering are used for text clustering, which
groups similar documents together based on their content. Text cluster‑
ing assists in organizing and summarizing massive document collections,
allowing for more effective information retrieval.
8. Deep Learning for Text: Deep learning models, particularly recurrent and
convolutional neural networks (CNNs), have performed admirably in a vari‑
ety of NLP applications. They can learn complicated patterns and hierarchi‑
cal representations from text input automatically, resulting in cutting‑edge
performance in tasks such as machine translation, text summarization, and
question answering.
9. Continuous Improvement: Machine‑learning models can be trained itera‑
tively, allowing them to improve over time as they encounter new data. This
adaptability is valuable in dynamic text mining scenarios, where data dis‑
tribution and characteristics may change.
10. Task Automation: Machine learning automates a variety of text mining
operations, decreasing manual labor and enhancing efficiency. This auto‑
mation is especially important when dealing with vast amounts of text data
that would be impossible to evaluate manually.
238 Data-Centric Artificial Intelligence for Multidisciplinary Applications

In summary, machine learning enables text mining by providing sophisticated tools

and algorithms capable of extracting valuable information, patterns, and insights
from unstructured text data, allowing businesses, researchers, and analysts to make
more informed decisions and gain a deeper understanding of the textual information
at their disposal.

16.1.3 Key Contributions

1. Holistic Overview: This chapter furnishes an all‑encompassing under‑
standing of machine‑learning methods specifically designed for text min‑
ing. It equips readers with a solid foundation in the subject’s essentials.
2. Practical Significance: The chapter underscores the practical importance
of text mining across diverse domains, from marketing to healthcare,
emphasizing its role in informed decision‑making.
3. Challenges and Opportunities: By outlining the challenges and potential
in text mining, this chapter prepares readers for navigating the intricacies of
the field.
4. Preprocessing Profundity: Emphasizing the role of preprocessing, the
chapter delineates techniques like normalization and NER, offering insights
into data enhancement.
5. Feature Extraction Focus: Through the elucidation of BoW, TF‑IDF, and
word embeddings, the chapter underscores the essence of feature extraction
for meaningful text analysis.
6. Supervised Learning Clarity: With a focus on techniques like Naive
Bayes, SVM, and decision trees, the chapter provides a clear path to under‑
standing supervised text classification.
7. Unveiling Unsupervised Techniques: Exploration of the K‑means, hierar‑
chical clustering, and LDA techniques demystifies unsupervised text clus‑
tering and topic modeling.
8. Sentiment and Opinion Discernment: Specific attention to sentiment
analysis, from lexicon‑based methods to deep learning models, equips read‑
ers with tools to comprehend public sentiment.
9. Advanced Applications: Through real‑world applications and advanced
techniques like text summarization, text generation, and question‑­answering
systems, readers gain insights into text mining’s diverse applications.5
10. Ethics and Future Prospects: Ethical considerations, privacy concerns,
and future trends are explored, shaping responsible and forward‑looking
text mining practices.

By encapsulating these contributions, this chapter empowers readers, be they

researchers, practitioners, or students, with a robust toolkit for harnessing machine
learning’s capabilities in text mining. It’s not merely a theoretical discourse but a
pragmatic guide to extract valuable insights from textual data and pave the way for
impactful decision‑making.
Machine‑Learning Techniques for Effective Text Mining 239

16.2 PREPROCESSING AND TEXT REPRESENTATION

Text mining requires preprocessing and text representation, which include translat‑
ing raw unstructured text input into a structured format suitable for analysis using
machine‑learning techniques.3 These stages aid in the reduction of noise, the extrac‑
tion of useful information, and the representation of textual data in a numerical for‑
mat that can be fed into machine‑learning models. Let’s take a closer look at each of
these steps:

16.2.1 Data Cleaning and Normalization

• Data cleaning involves removing any irrelevant characters, symbols, or spe‑
cial characters from the text data. It may also include converting all text to
lowercase to ensure case‑insensitive analysis.
• Normalization aims to standardize the text data by converting contractions
to their full forms (e.g., “can’t” to “cannot”) and expanding abbreviations
(e.g., “USA” to “United States of America”). This step helps in reducing
lexical variations and improves the consistency of the data.3

16.2.2 Tokenization and Stemming

• Tokenization is the process of dividing a text into smaller units called
tokens, which can be individual words or even smaller units like n‑grams
(n‑word sequences). Tokenization makes feature extraction and representa‑
tion easier in subsequent steps.
• The process of reducing words to their base or root form by removing suf‑
fixes or prefixes is known as stemming. For example, stemming “running,”
“runs,” and “run” yields the root word “run.” This phase reduces the dimen‑
sionality of the data while increasing computational efficiency.

16.2.3 Stopword Removal
• Stopwords are common words that appear frequently in the language (e.g.,
“the,” “and,” and “is”) but rarely add significantly to the overall meaning of
the text. Stopwords are removed to reduce noise and focus on more relevant
words that provide crucial information.

16.2.4 Feature Extraction Methods for Text Data

• Bag‑of‑Words (BoW): A popular text representation technique, BoW turns
text into a numerical vector that represents the frequency of each word in
the document. Each word is processed individually, and the resulting vector
is known as a term frequency (TF) vector.
• TF‑IDF (Term Frequency‑Inverse Document Frequency): TF‑IDF is a
weighting technique that gives more weight to words that appear frequently
240 Data-Centric Artificial Intelligence for Multidisciplinary Applications

in a document but are uncommon throughout the corpus. It aids in high‑

lighting key and distinguishing words in each manuscript.5
• Word embeddings are dense vector representations of words in a continuous
vector space that capture semantic links between words. Word embeddings
that have been pretrained, such as Word2Vec or GloVe, can be used to rep‑
resent words in a more meaningful way.

16.2.5 Handling Sparse Data and Dimensionality Reduction Techniques

• Because of the wide vocabulary, text data frequently yields sparse
high‑dimensional feature vectors. Dimensionality reduction techniques
such as Singular Value Decomposition or Principal Component Analysis
(PCA) can be used to minimize the feature space while keeping vital infor‑
mation to deal with this.6
• Furthermore, approaches such as feature selection or feature engineering
can be used to focus on the most important traits while removing noisy or
irrelevant ones.

Preprocessing and text representation help to turn raw text data into an organized
and comprehensible manner, laying the groundwork for effective text mining.
These transformed representations can then be fed into various machine‑­
learning algorithms for tasks such as text classification, clustering, sentiment
analysis, topic modeling, and more. Proper preprocessing and representation
are essential for extracting accurate and useful insights from text data as well as
increasing the overall performance of text mining models. Figure 16.2 depicts
the activity diagram detailing the preprocessing procedures within the context
of text mining.

16.3 SUPERVISED LEARNING FOR TEXT CLASSIFICATION

Supervised learning for text classification is a machine‑learning approach where a
model is trained on labeled text data to predict the category or class of new, unseen
text instances.5 It involves using a set of input features (representations of text data)
and their corresponding labels (class labels) to train a classifier that can make predic‑
tions on new, unlabeled text instances. Text classification has numerous real‑world
applications, including spam detection, sentiment analysis, topic categorization, and
document classification.

16.3.1 SUPERVISED LEARNING ALGORITHMS

FOR TEXT CLASSIFICATION
1. Naive Bayes:
• Naive Bayes is a probabilistic classification technique that is based on
the theorem of Bayes. Despite its simplicity, it has performed admirably
in a variety of text classification tasks.
Machine‑Learning Techniques for Effective Text Mining 241

FIGURE 16.2 Preprocessing working.

242 Data-Centric Artificial Intelligence for Multidisciplinary Applications

• In the context of text categorization, Naive Bayes posits that the pres‑
ence of one word in a document is unrelated to the presence of other
words (thus the term “naive”). It computes the likelihood of a text
belonging to a specific class based on the likelihood of individual words
occurring in that class.
• Naive Bayes is a computationally efficient method that requires little
training data. It is effective for high‑dimensional feature spaces such as
BoW representations.
2. Support Vector Machines (SVMs):
• SVMs are a strong and commonly used supervised learning technique
for text classification. Its goal is to determine the ideal hyperplane in a
high‑dimensional space for separating data points of different classes.
• SVM attempts to discover the hyperplane in the feature space that opti‑
mizes the margin between data points of various classes in the context
of text classification. It is useful for applications with complex decision
boundaries and can efficiently handle high‑dimensional feature vectors.
• SVM can handle noisy data and performs well with less training data.
However, when dealing with very huge datasets, it may experience scal‑
ing challenges.
3. Decision Trees:
• Decision Trees are a non‑parametric supervised learning approach that
can be used for classification and regression applications. They partition
the data recursively based on feature values to generate a tree‑like struc‑
ture that predicts the class labels of subsequent occurrences.
• Decision trees in text classification make binary judgments at each node
based on the values of specific words or features, which results in the
assignment of a class label at the tree’s leaves.
• Decision Trees are easily interpretable and visualized, making them
useful for understanding the decision‑making process in text categori‑
zation problems. They are, nevertheless, prone to overfitting, particu‑
larly with deep trees and noisy data.
4. Neural Networks:
• Because of their ability to automatically learn complicated patterns
from text input, neural networks, particularly deep learning models,
have gained great interest in text categorization.
• For sequential data such as words or texts, RNNs and long short‑term
memory (LSTM) networks are appropriate. They are capable of captur‑
ing text’s sequential dependencies, making them useful for sentiment
analysis and language modeling.
• CNNs are frequently employed for text classification problems, par‑
ticularly when dealing with fixed‑length input like BoW representa‑
tions. From text data, CNNs can learn local patterns and hierarchical
representations.
• Transfer learning with pretrained language models such as BERT and
GPT‑3 has also demonstrated exceptional performance in text categori‑
zation tasks, exploiting knowledge obtained from large‑scale pretraining.
Machine‑Learning Techniques for Effective Text Mining 243

Each supervised learning algorithm has its strengths and weaknesses when applied
to text classification. The choice of algorithm depends on factors like the size of
the dataset, the complexity of the task, the interpretability required, and the avail‑
ability of computational resources. Proper evaluation and experimentation are
crucial to selecting the most suitable algorithm for a specific text classification prob‑
lem. Figure 16.3 presents the activity diagram delineating the operational flow of
Supervised Learning for Text Classification within the domain of text mining.

16.3.2 Feature Engineering for Text Classification Tasks

Text classification feature engineering is translating raw text data into meaning‑
ful and useful numerical representations (features) that may be utilized as input to
machine‑learning algorithms. By capturing essential information and patterns from
text data, effective feature engineering plays a crucial role in increasing the per‑
formance of text classification models. Here are some examples of commonly used
feature engineering techniques in text classification:

1. Bag‑of‑Words (BoW):
• BoW is a common and simple text data representation approach.
It entails generating a vocabulary of unique words (or tokens) found
across the dataset. Each document is then represented numerically as a
vector, with each entry representing the frequency of a certain term in
the document.
• BoW treats each word separately and disregards word order and con‑
text. While straightforward, it captures the incidence of various words
in documents, making it appropriate for text categorization tasks such
as sentiment analysis and spam detection.
2. TF‑IDF (Term Frequency‑Inverse Document Frequency):
• The TF‑IDF representation is a variant of the BoW representation that
assigns a weight to each word in the document to indicate its relevance
in the document relative to the overall dataset.
• TF is a measure of the frequency of a term in a document, whereas
Inverse Document Frequency (IDF) is a measure of the rarity of a word
across all documents. The TF‑IDF weight is the product of these two
variables.
• TF‑IDF aids in the identification of words that are discriminative and
informative for a given class because they are common in the document
of interest but uncommon in other documents.
3. Word Embeddings:
• Word embeddings are continuous vector space‑dense vector representa‑
tions of words. Word embeddings represent the semantic relationships
between words, allowing algorithms to comprehend word similarity
and context.
• In text classification tasks, pretrained word embeddings such as
Word2Vec, GloVe, and FastText are extensively employed. These
embeddings are derived from big text corpora and can be applied to
244 Data-Centric Artificial Intelligence for Multidisciplinary Applications
Supervised learning for text classification.
FIGURE 16.3
Machine‑Learning Techniques for Effective Text Mining 245

new text classification tasks to boost performance, particularly when

training data is scarce.
4. Part‑of‑Speech (POS) Tagging:
• POS tagging entails marking each word in a document with the appro‑
priate part of speech (e.g., noun, verb, adjective). POS tags can provide
useful information about sentence grammaticality and can be employed
as features in some text categorization tasks.
5. N‑Grams:
• N‑grams are contiguous sequences of N words taken from a single doc‑
ument. Using N‑grams as features can assist capture local word order
and context, giving the classifier extra information.
• Unigrams (N = 1) represent single words, but bigrams (N = 2) and
trigrams (N = 3) represent two‑word and three‑word sequences,
respectively.
6. Domain‑Specific Features:
• Domain‑specific traits can be important and informative in some text
categorization tasks. In sentiment analysis of customer evaluations, for
example, features such as the presence of emoticons, capitalization, or
special characters may contain sentiment‑related information.
7. Word Frequency and Document Length:
• As features, basic statistics such as word frequency and document
length (the amount of words in a document) can be employed. These
data may provide insight into the text’s writing style or subject.

It is vital to note that the selection of feature engineering techniques is depen‑

dent on the individual properties of the text data as well as the classification task
requirements. Feature engineering is an iterative process, and experimenting with
various strategies might assist in identifying the most relevant features for a cer‑
tain text classification task. Combining multiple feature engineering strategies
can also result in increased model performance in complicated text classification
tasks.

16.3.3 Case Studies and Real‑World Examples of Text Classification Tasks

16.3.3.1 Case Study 1: Sentiment Analysis in Social Media
Objective: To perform sentiment analysis on social media data to understand the
overall sentiment (positive, negative, or neutral) of tweets related to a specific prod‑
uct or brand.
Data: A large collection of tweets mentioning the product/brand, manually
labeled as positive, negative, or neutral sentiment.
Approach:

1. Data Preprocessing: Clean the text data and remove URLs, special char‑
acters, and emojis. Tokenize the tweets and remove stopwords.
2. Feature Engineering: Convert the preprocessed tweets into numerical rep‑
resentations using TF‑IDF or word embeddings.
246 Data-Centric Artificial Intelligence for Multidisciplinary Applications

3. Model Selection: Train numerous classifiers such as Naive Bayes, SVMs,

and deep learning models such as LSTM or BERT.
4. Model Evaluation: To evaluate the performance of each model, use metrics
such as accuracy, precision, recall, and F1 score.
5. Prediction: Deploy the best‑performing model to predict sentiment for new
tweets in real time.

Real‑World Example: A company wants to gauge public sentiment about their latest
product release. They use sentiment analysis to analyze thousands of tweets men‑
tioning the product. The analysis reveals that overall sentiment is positive, but some
negative feedback points to specific issues that need to be addressed for product
improvement.

16.3.3.2 Case Study 2: Spam Detection in Emails

Objective: To build a classifier that can automatically identify spam emails and dis‑
tinguish them from legitimate ones.
Data: A labeled dataset containing emails, marked as either spam or non‑spam
(ham).
Approach:

1. Data Preprocessing: Clean the emails, remove HTML tags, and normalize
the text (e.g., convert to lowercase).
2. Feature Engineering: Convert the text into numerical representations
using BoW or TF‑IDF.
3. Model Selection: Train classifiers like Naive Bayes, SVMs, or Decision
Trees.
4. Model Evaluation: Use metrics like accuracy, precision, recall, and F1
score to evaluate the model’s performance on a separate test set.
5. Prediction: Deploy the best‑performing model to automatically classify
incoming emails as spam or ham.

Real‑World Example: An email service provider wants to protect its users from
spam emails. By employing a spam detection system using supervised learning, they
can accurately filter spam emails and improve user experience and security.

16.4 UNSUPERVISED LEARNING FOR TEXT CLUSTERING

Unsupervised learning for text clustering is a machine‑learning approach where the
goal is to automatically group similar text documents into clusters without using any
predefined class labels. Text clustering aims to discover hidden patterns and struc‑
tures within the text data, allowing for effective organization and summarization of
large document collections.5
Approach:

1. Data Preprocessing: Similar to supervised learning, the text data under‑

goes preprocessing steps like data cleaning, normalization, tokenization,
stemming, and stopword removal.
Machine‑Learning Techniques for Effective Text Mining 247

2. Text Representation: The preprocessed text data is then converted into

numerical representations using techniques like BoW, TF‑IDF, or word
embeddings. These representations capture the semantic meaning and con‑
text of words, making it possible to compute similarities between documents.
3. Clustering Algorithms: Various unsupervised clustering algorithms are
used to group similar documents together. Some commonly used algorithms
include K‑means, Hierarchical Clustering, DBSCAN, and Mean Shift.
4. Similarity Metrics: To measure the similarity between documents, appro‑
priate distance or similarity metrics like Cosine Similarity or Euclidean
Distance are employed. These metrics help determine how close or dissimi‑
lar documents are in the feature space.
5. Clustering Evaluation: Unlike supervised learning, there are no pre‑
defined class labels for unsupervised text clustering. As a result, evaluation
becomes more challenging. Internal clustering evaluation metrics, such as
the Silhouette Score or the Davies–Bouldin Index, can be used to assess the
quality of clustering results.
6. Cluster Interpretation: Once the clustering is performed, post‑processing
steps may be employed to interpret and label the clusters based on their con‑
tent or characteristics. This step is important for understanding the topics or
themes represented by each cluster.

Benefits:

1. Unsupervised text clustering is valuable in exploratory data analysis and

knowledge discovery, especially when there is no prior knowledge of the
document categories.
2. It helps in organizing large document collections, enabling more efficient
information retrieval and content recommendation systems.
3. Unsupervised clustering can be used as a preprocessing step for supervised
learning tasks, reducing the dimensionality and complexity of the text data.

Challenges:

1. Determining the optimal number of clusters (K) can be challenging, as

there is no ground truth available for unsupervised text clustering tasks.
4. Text data can be high‑dimensional and sparse, making it challenging to find
meaningful clusters and handle noise.
5. The effectiveness of clustering heavily depends on the choice of text repre‑
sentation and clustering algorithm, which requires careful consideration.5

Unsupervised learning for text clustering is a valuable technique for discovering pat‑
terns and organizing large text corpora without the need for explicit class labels. Its
applications span across various domains, including information retrieval, content
analysis, and document organization. However, proper evaluation and interpretation
of clusters are essential for deriving meaningful insights from unsupervised text clus‑
tering. Figure 16.4 delineates the operational mechanics underlying Unsupervised
Learning for Text Clustering.
248 Data-Centric Artificial Intelligence for Multidisciplinary Applications

FIGURE 16.4 Unsupervised learning for text clustering.

16.4.1 Unsupervised Clustering Algorithms

1. K‑means Clustering:
• K‑means is a widely used partition‑based clustering algorithm that aims
to divide data points into K clusters.
• It starts by randomly initializing K cluster centroids in the feature
space. Each data point is then assigned to the nearest centroid based on
a distance metric (usually Euclidean distance).
• After the initial assignments, the centroids are updated to the mean of
the data points within each cluster. The assignment and centroid update
steps are repeated iteratively until convergence.
• K‑means is simple, fast, and effective for spherical‑shaped clusters.
However, it requires specifying the number of clusters (K) beforehand,
and its performance may degrade with non‑linear or irregularly shaped
clusters.
2. Hierarchical Clustering:
• Hierarchical clustering creates a hierarchy of clusters by iteratively
merging or splitting clusters based on similarity.
• It can be performed in two ways: Agglomerative (bottom‑up) and
Divisive (top‑down).
• Agglomerative starts with each data point as its own cluster and repeat‑
edly merges the most similar clusters until a single cluster containing
all data points is formed.
• Divisive starts with all data points in one cluster and repeatedly splits
the least similar clusters until each data point is in its own cluster.
• Hierarchical clustering produces a dendrogram, which can be used to
visualize the clustering hierarchy. The number of clusters is determined
by cutting the dendrogram at an appropriate level.
• Hierarchical clustering is useful for understanding the hierarchical
structure of data and is not sensitive to the number of clusters.
Machine‑Learning Techniques for Effective Text Mining 249

3. Density‑Based Spatial Clustering of Applications with Noise (DBSCAN):

• DBSCAN is a density‑based clustering algorithm that groups data
points based on their density and identifies noise points as well.
• It defines two important parameters: Epsilon (ε) and MinPts. Epsilon
determines the neighborhood radius around each data point, and
MinPts sets the minimum number of points required to form a dense
region (core point).
• A core point is a data point with at least MinPts neighbors within its
ε‑radius. Any data point within the ε‑radius of a core point is considered
part of the same cluster. Non‑core points that are within the ε‑radius of
a core point are called border points and may belong to the same cluster
or be considered noise.
• DBSCAN is robust to noise and can discover clusters of arbitrary
shapes. It does not require specifying the number of clusters beforehand
and is suitable for datasets with varying cluster densities.

Each of these unsupervised clustering algorithms has its strengths and weaknesses.
The choice of algorithm depends on the specific characteristics of the data and the
nature of the clusters to be discovered. Experimenting with different algorithms is
essential to find the most suitable one for a particular clustering task.

16.4.2 FEATURE REPRESENTATION FOR CLUSTERING TASKS

Feature representation for clustering tasks involves transforming raw data into a
suitable numerical format that captures the essential characteristics of the data for
clustering.6 Effective feature representation is crucial for achieving meaningful and
accurate clustering results. Here are some common feature representation techniques
for clustering tasks:

1. Numeric Feature Representation: For data with numerical attributes, no

special preprocessing is required, and the features can be directly used as
they are.
2. One‑Hot Encoding: For categorical data, one‑hot encoding is used to con‑
vert categorical variables into a binary representation. Each category is rep‑
resented as a binary vector, where only one element is set to 1, indicating the
presence of that category.
3. Bag‑of‑Words (BoW): BoW is a widely used representation for text data
in clustering tasks. It converts text documents into numerical vectors repre‑
senting the frequency of each word in the document. Each word is treated
as a separate feature, and the resulting vector is called a term frequency
vector.
4. Term Frequency‑Inverse Document Frequency (TF‑IDF): TF‑IDF is an
extension of BoW that assigns higher weights to words that are frequent in
a document but rare across the entire corpus. It helps in highlighting impor‑
tant and discriminative words for each document.
250 Data-Centric Artificial Intelligence for Multidisciplinary Applications

5. Word Embeddings: Word embeddings are dense vector representations of

words in a continuous vector space. They capture the semantic relationships
between words, making them effective for text clustering tasks.
6. Image Feature Extraction: For image data, feature extraction techniques
like Histogram of Oriented Gradients, Local Binary Patterns, or pretrained
deep learning models (e.g., CNN features) can be used to obtain numerical
representations of images.
7. Vector Space Model (VSM): VSM represents documents as vectors in a
multi‑dimensional space, where each dimension corresponds to a term in
the document collection. The presence or absence of terms in each docu‑
ment determines the vector’s value.
8. Graph‑Based Feature Representation: For data with a graph structure
(e.g., social networks), graph‑based representations like node embeddings
or graph kernels can be used to convert the graph data into numerical vec‑
tors suitable for clustering.
9. Statistical Features: Statistical measures like mean, standard deviation,
or variance can be used to represent data points based on their attribute
values.5
10. Composite Features: Combining multiple features or applying dimension‑
ality reduction techniques like PCA or t‑distributed Stochastic Neighbor
Embedding (t‑SNE) can be beneficial for capturing complex relationships
and reducing the feature space’s dimensionality.

The choice of feature representation depends on the nature of the data and the clus‑
tering task at hand. Proper feature engineering is essential for achieving accurate
and meaningful clustering results and plays a critical role in the success of clustering
algorithms.

16.4.3 Case Studies Illustrating Text Clustering Techniques

16.4.3.1 Case Study 1: News Article Clustering
Objective: Cluster a collection of news articles into distinct groups based on their
topics and content.
Data: A dataset of news articles from various sources, such as newspapers, blogs,
and online media. Each article contains the article text and its publication source.
Approach:

1. Data Preprocessing: Clean the text data by removing special characters,

numbers, and stopwords. Tokenize the text into words and apply stemming
or lemmatization to reduce words to their base form.
2. Feature Representation: Represent the preprocessed articles using TF‑IDF
or word embeddings to obtain numerical vectors for each article. TF‑IDF
captures the importance of words in each document relative to the entire cor‑
pus, while word embeddings represent words in a continuous vector space.
Machine‑Learning Techniques for Effective Text Mining 251

3. Clustering Algorithm: Apply K‑means clustering algorithm to group simi‑

lar articles together. Choose the optimal number of clusters (K) using tech‑
niques like the Elbow Method or Silhouette Score.
4. Cluster Evaluation: Evaluate the clustering results using internal evalu‑
ation metrics like Silhouette Score or external evaluation metrics like
Adjusted Rand Index (ARI) if ground truth labels are available.
5. Visualization: Visualize the clustering results using dimensionality reduc‑
tion techniques like t‑SNE to plot the articles in a 2D space based on their
numerical representations. Assign each article a color corresponding to its
cluster label.

Real‑World Application: This clustering technique can be used by news aggrega‑

tor platforms to categorize and organize news articles into different topics, enabling
users to access relevant news easily and efficiently.

16.4.3.2 Case Study 2: Customer Review Clustering

Objective: Group customer reviews of a product into distinct clusters based on the
sentiment and topics expressed in the reviews.
Data: A dataset of customer reviews for a specific product, containing the review
text and corresponding ratings (e.g., 1–5 stars).
Approach:

1. Data Preprocessing: Clean the review text by removing special charac‑

ters, punctuation, and stopwords. Perform sentiment analysis to categorize
reviews as positive, negative, or neutral based on their star ratings.
2. Feature Representation: Use TF‑IDF or word embeddings to represent the
preprocessed review text as numerical vectors. Additionally, consider add‑
ing sentiment scores as additional features.
3. Clustering Algorithm: Apply K‑means or DBSCAN clustering algorithm
to group reviews with similar sentiments and topics together. Experiment
with different sentiment‑based and content‑based features.
4. Cluster Evaluation: Evaluate the clustering results using internal evalua‑
tion metrics like Silhouette Score or external evaluation metrics like ARI if
ground truth sentiment labels are available.
5. Interpretation: Analyze the reviews within each cluster to understand the
common sentiments and topics expressed by customers.

Real‑World Application: Businesses can use this clustering technique to gain

insights from customer feedback, identify common issues or strengths in their
products, and make data‑driven decisions for product improvements or marketing
strategies.
These case studies demonstrate how text clustering techniques can be applied to
real‑world scenarios, such as organizing news articles into topics and understanding
customer sentiments from reviews. Text clustering is a powerful tool for knowledge
discovery, information retrieval, and customer insights in various industries.
252 Data-Centric Artificial Intelligence for Multidisciplinary Applications

16.5 SENTIMENT ANALYSIS AND OPINION MINING

16.5.1 Sentiment Analysis and its Significance
Sentiment Analysis, also known as opinion mining, is an NLP technique that aims
to identify and extract subjective information from text data, especially determining
the sentiment or emotion expressed in the text. The goal of sentiment analysis is to
classify the text as having positive, negative, neutral, or sometimes more nuanced
sentiments like happiness, sadness, anger, etc.1
Approach:

1. Data Preprocessing: Clean the text data and remove irrelevant informa‑
tion, special characters, and punctuation. Tokenize the text into words and
apply stemming or lemmatization to reduce words to their base form.
2. Feature Representation: Convert the preprocessed text into numerical
representations using techniques like BoW, TF‑IDF, or word embeddings.
These representations capture the semantic meaning and context of words.
3. Sentiment Classification: Utilize supervised machine‑learning algorithms,
such as Naive Bayes, SVMs, or deep learning models like LSTM or BERT,
to classify the sentiment of the text. These models are trained on labeled
data with sentiment annotations.3
4. Sentiment Analysis Output: The output of sentiment analysis is the clas‑
sification of the text into positive, negative, or neutral sentiments.

Applications:

1. Business and Product Reviews: Sentiment analysis is used to analyze

customer reviews and feedback to understand the overall sentiment toward
products or services.
2. Social Media Monitoring: Companies monitor social media platforms to
gauge public sentiment toward their brand or marketing campaigns.
3. Market Research: Sentiment analysis helps businesses in market research
to identify trends, customer preferences, and brand perceptions.
4. Customer Support: Sentiment analysis can be applied to customer support
interactions to gauge customer satisfaction and detect potential issues.5

Significance:

1. Customer Insights: In business, sentiment analysis is crucial for under‑

standing customer sentiment toward products, services, and brands. It helps
companies gauge customer satisfaction, identify areas for improvement, and
make data‑driven decisions for marketing and product development.
2. Brand Reputation Management: Sentiment analysis aids in monitor‑
ing and managing brand reputation on social media and other platforms.
Companies can address negative sentiment and respond to customer feed‑
back promptly.
Machine‑Learning Techniques for Effective Text Mining 253

3. Market Research: Sentiment analysis is used in market research to analyze

public opinions on products, competitors, and market trends. It provides
valuable insights for businesses to develop competitive strategies.
4. Social Media Analysis: Sentiment analysis is extensively used in social media
monitoring to understand public sentiment toward specific topics, events, or
public figures. It helps in gauging public reactions and predicting trends.
5. Customer Service and Support: Sentiment analysis is employed in cus‑
tomer service and support interactions to assess customer satisfaction levels
and identify potential issues or dissatisfied customers.
6. Political Analysis: In political contexts, sentiment analysis is used to ana‑
lyze public sentiment toward political figures, policies, and events. It can
provide insights into public opinion during election campaigns.
7. Product Reviews: Sentiment analysis assists in analyzing and summariz‑
ing product reviews and feedback to identify recurring themes and senti‑
ments among customers.

Overall, sentiment analysis is a powerful tool for understanding human emotions

and opinions from vast amounts of text data. Its significance extends across various
domains, helping businesses, researchers, and organizations make informed deci‑
sions, enhance customer experiences, and improve overall communication strate‑
gies.4 Illustrated in Figure 16.5 is the procedural representation of the operational
workflow within the domain of sentiment analysis.

16.5.2 Supervised and Unsupervised Approaches to Sentiment Analysis

1. Supervised Sentiment Analysis: Supervised sentiment analysis is a
machine‑learning approach where the sentiment analysis model is trained
on a labeled dataset that contains text samples along with their correspond‑
ing sentiment labels (positive, negative, neutral, etc.). The model learns
from this labeled data to make predictions on new, unseen text data.
• Approach:
– Data Preparation: The labeled dataset is prepared with text
samples and sentiment labels. Text data undergoes preprocessing,
including cleaning, tokenization, and normalization.
– Feature Representation: The preprocessed text is converted into
numerical representations using techniques like BoW, TF‑IDF, or
word embeddings.1
– Model Training: Supervised learning algorithms like Naive Bayes,
SVMs, Logistic Regression, or deep learning models like LSTM
or BERT are trained on the labeled data to learn the relationships
between the textual features and sentiment labels.
– Model Evaluation: The trained model is evaluated on a separate
test dataset to measure its performance using metrics like accuracy,
precision, recall, F1 score, etc.
– Prediction: Once the model is trained and evaluated, it can be used
to predict the sentiment of new, unseen text data.
254 Data-Centric Artificial Intelligence for Multidisciplinary Applications

FIGURE 16.5 Sentiment analysis working.

Machine‑Learning Techniques for Effective Text Mining 255

• Pros:
– Supervised learning provides accurate sentiment predictions when
trained on sufficient and representative labeled data.
– It can handle complex relationships between features and sentiment
labels.
• Cons:
– Requires a large amount of labeled data for training.
– Performance may suffer if the training data is biased or does not
fully represent the distribution of sentiments in real‑world data.
2. Unsupervised Sentiment Analysis: Unsupervised sentiment analysis is an
approach where the sentiment analysis model is not provided with labeled
data during training. Instead, it aims to identify patterns and structures in
the data without predefined sentiment labels
• Approach:
– Data Preparation: Text data undergoes preprocessing, including
cleaning, tokenization, and normalization.
– Feature Representation: The preprocessed text is converted into
numerical representations using techniques like BoW, TF‑IDF, or
word embeddings.
– Clustering: Unsupervised clustering algorithms like K‑means,
Hierarchical Clustering, or DBSCAN are applied to group similar
text samples together based on their numerical representations.
– Sentiment Assignment: Sentiments are assigned to the clusters
based on the predominant sentiment of the text samples within each
cluster. For example, if most samples in a cluster are positive, that
cluster is assigned a positive sentiment.
– Evaluation (Optional): Since unsupervised sentiment analysis
doesn’t have labeled data for evaluation, the quality of the clusters
can be assessed using internal clustering evaluation metrics.
• Pros:
– Unsupervised approaches can be applied when labeled data is
scarce or unavailable.
– It can discover hidden patterns and structures in the data without
relying on predefined sentiment labels.
• Cons:
– The sentiment assignments may not always match human judgment
or predefined sentiment labels.
– It can be challenging to interpret and validate the accuracy of the
results in the absence of labeled data.

Hybrid Approaches: Some sentiment analysis systems combine both supervised

and unsupervised techniques. For example, an unsupervised approach might be
used to cluster data into groups representing different sentiments, and then a smaller
labeled dataset might be used to train a supervised model on each cluster to fine‑tune
sentiment predictions.
256 Data-Centric Artificial Intelligence for Multidisciplinary Applications

Overall, the choice between supervised and unsupervised approaches depends

on the availability of labeled data, the nature of the sentiment analysis task, and the
desired accuracy and interpretability of the results.

16.5.3 Applications of Sentiment Analysis in Social Media

1. Brand Monitoring: Companies use sentiment analysis to monitor mentions
of their brand on social media platforms. It helps them gauge public senti‑
ment and identify potential issues or negative feedback that require prompt
responses.
2. Social Listening: Sentiment analysis allows businesses to listen to what
customers are saying about their products, services, or marketing cam‑
paigns. It provides insights into customer preferences and expectations.
3. Reputation Management: Brands use sentiment analysis to manage their
online reputation. It helps in identifying and addressing negative sentiment
and managing public perception.
4. Campaign Evaluation: Sentiment analysis is employed to evaluate the suc‑
cess of marketing campaigns or product launches. Positive sentiment indi‑
cates a well‑received campaign, while negative sentiment indicates areas for
improvement.
5. Customer Support: Sentiment analysis assists in customer support interac‑
tions on social media. It helps prioritize and address customer issues based
on sentiment.
6. Influencer Marketing: Brands use sentiment analysis to assess the senti‑
ment of influencer content and its impact on the brand’s reputation.

Applications of Sentiment Analysis in Customer Feedback:

1. Product and Service Improvements: Sentiment analysis of customer feed‑

back helps companies identify specific product features or service aspects
that need improvement based on customer sentiments.
2. Customer Satisfaction Measurement: Sentiment analysis allows busi‑
nesses to measure overall customer satisfaction by analyzing sentiments in
feedback.5
3. Competitor Analysis: Sentiment analysis helps compare customer senti‑
ments toward competitors’ products and services, providing insights for
competitive strategy.
4. Market Research: Sentiment analysis is used in market research to gather
consumer opinions and preferences, helping businesses develop targeted
marketing strategies.5
5. Review Aggregation: Sentiment analysis is used in review aggregators to
classify reviews as positive, negative, or neutral, providing a summary of
overall customer sentiments.
6. Feedback Trend Analysis: Sentiment analysis assists in analyzing trends
in customer feedback over time, identifying changing sentiments or emerg‑
ing issues.
Machine‑Learning Techniques for Effective Text Mining 257

Overall, sentiment analysis plays a significant role in understanding customer senti‑

ments, preferences, and opinions. It provides businesses with valuable insights for
making informed decisions, enhancing customer experiences, and improving prod‑
ucts and services based on customer feedback.

16.6 NER AND ENTITY LINKING

16.6.1 Introduction to NER
NER is an NLP task that involves identifying and classifying named entities in a
text into predefined categories such as person names, organization names, location
names, date expressions, numerical quantities, and more.1 Named entities are spe‑
cific words or phrases that refer to unique entities in the real world, such as names of
people, places, companies, products, and dates.1
Objective:
The main goal of NER is to automatically recognize and classify these named
entities in a given text, extracting valuable information and structuring unstructured
text data into a more structured format.
Approach:

1. Data Preprocessing: The text data is cleaned and tokenized into individual
words or phrases.
2. Linguistic Features: NER systems often use linguistic features, such as
POS tagging, to identify named entities. Certain POS patterns are indicative
of named entities, like proper nouns.
3. Machine‑learning Models: NER is often approached as a supervised
learning problem, where machine‑learning models are trained on annotated
text data that includes the labeled named entities. Popular machine‑learning
algorithms like conditional random fields (CRFs), SVMs, or deep learning
models like bidirectional long short‑term memory (BiLSTM) networks are
commonly used.2
4. Named Entity Classification: During training, the model learns to classify
each word or phrase in the text into predefined categories, such as “Person,”
“Organization,” “Location,” etc.
5. Named Entity Extraction: Once the model is trained, it can be used to pro‑
cess new, unseen text data and identify and extract named entities present in
the text.

Significance:
• NER has various practical applications across different domains:
– Information Extraction: NER is used to extract valuable informa‑
tion from unstructured text, helping in knowledge discovery and
information retrieval.
– Search Engines: NER improves the accuracy of search engines by
identifying and recognizing entities mentioned in search queries or
web pages.
258 Data-Centric Artificial Intelligence for Multidisciplinary Applications

– Question Answering: NER aids in identifying relevant entities

in the context of a question, enabling better responses in ques‑
tion‑answering systems.
– Chatbots and Virtual Assistants: NER helps chatbots and virtual
assistants understand user queries and provide more accurate and
contextually relevant responses.
– Document Categorization: NER assists in categorizing docu‑
ments based on the named entities mentioned, aiding in content
organization.
– Information Retrieval: NER enhances the precision and recall
of information retrieval systems by indexing and linking named
entities.

Overall, NER is a fundamental NLP task that plays a crucial role in various down‑
stream applications, improving the understanding and processing of text data by
identifying and classifying specific entities of interest. Depicted in Figure 16.6 is the
sequential process illustrating the operational flow of NER functioning.

16.6.2 NER Techniques: Rule‑Based and Machine‑learning Methods

1. Rule‑Based NER Techniques: Rule‑based NER techniques rely on pre‑
defined rules or patterns to identify and classify named entities in text
data. These rules are usually created manually by domain experts or lin‑
guists based on linguistic patterns and characteristics of named entities.1
Rule‑based approaches are straightforward to implement and can be useful
when dealing with specific types of entities or domains where labeled train‑
ing data may be limited.
• Approach:
– Rule Definition: Linguists or domain experts define rules based on
patterns, regular expressions, or syntactic structures that are indica‑
tive of named entities. For example, patterns like capitalization,
specific POS tags, or word sequences can be used to identify person
names, locations, or organizations.
– Tokenization: The text data is tokenized into individual words or
phrases.
– Rule Application: The defined rules are applied to the tokenized
text to identify and classify named entities based on the specified
patterns.
– Entity Categorization: The recognized named entities are catego‑
rized into predefined entity types, such as “Person,” “Organization,”
“Location,” etc.
• Pros:
– Rule‑based NER is interpretable and allows experts to fine‑tune the
system easily.
– It can work well in specific domains or with limited training data.
Machine‑Learning Techniques for Effective Text Mining 259

FIGURE 16.6 NER working.

260 Data-Centric Artificial Intelligence for Multidisciplinary Applications

• Cons:
– Creating accurate and comprehensive rules can be labor intensive
and time‑consuming.
– Rule‑based approaches may struggle with handling complex or
ambiguous cases.
2. Machine‑learning‑Based NER Techniques: Machine‑learning‑based
NER techniques use supervised or unsupervised learning methods to auto‑
matically learn patterns and features from labeled training data. These
methods have become more popular due to their ability to handle complex
linguistic patterns and generalize well across different domains.
• Approach:
– Data Preparation: A labeled dataset is prepared, where the text
data is annotated with entity labels (e.g., “Person,” “Organization,”
“Location,” etc.).
– Feature Extraction: Features are extracted from the text data, such
as word embeddings, POS tags, contextual information, etc., to rep‑
resent the words in a numerical format suitable for machine‑learn‑
ing algorithms.
– Model Training: Supervised machine‑learning algorithms, like
CRF, SVMs, or deep learning models like BiLSTM networks, are
trained on the labeled data to learn the relationship between the
features and entity labels.
– Model Evaluation: The trained model is evaluated on a separate
test dataset to measure its performance using metrics like precision,
recall, F1 score, etc.
– Prediction: Once the model is trained and evaluated, it can be used
to predict named entities in new, unseen text data.
• Pros:
– Machine‑learning‑based NER can automatically learn complex
patterns and generalize to different contexts and domains.
– It can handle large amounts of data and adapt to new data.
• Cons:
– Requires a significant amount of labeled data for training.
– Model complexity and training time may be higher compared to
rule‑based methods.

Hybrid Approaches:
In practice, hybrid approaches that combine both rule‑based and machine‑learn‑
ing techniques are often used. Rule‑based methods can be used for specific entity
types or known patterns, while machine‑learning models can be applied to handle
more ambiguous cases or new entity types. Such hybrid approaches leverage the
strengths of both techniques to improve NER performance.
In conclusion, NER is a crucial task in NLP, enabling the extraction of valuable
information from text data. Both rule‑based and machine‑learning‑based techniques
have their advantages and applications, and the choice between them depends on the
specific requirements and characteristics of the NER task at hand.
Machine‑Learning Techniques for Effective Text Mining 261

16.6.3 Real‑World Applications of NER

1. Information Extraction: NER is widely used in information extraction
tasks to identify and extract specific pieces of information from unstruc‑
tured text. It helps in extracting names of people, organizations, loca‑
tions, dates, and more, making it easier to organize and analyze textual
data.
2. Search Engines: NER enhances search engine capabilities by recognizing
and categorizing named entities in search queries and web documents. It
improves search result relevance and provides more accurate answers to
user queries.
3. Social Media Monitoring: NER is used in social media monitoring tools
to identify and track mentions of specific entities, products, or brands. This
enables businesses to gain insights into social media conversations and pub‑
lic sentiments.
4. Customer Support: NER assists in automating customer support interac‑
tions by identifying entities mentioned in customer queries, allowing for
more personalized and efficient responses.
5. Chatbots and Virtual Assistants: NER is employed in chatbots and vir‑
tual assistants to understand and respond to user queries more accurately
and contextually.
6. Question–Answering Systems: NER plays a vital role in question–answer‑
ing systems by identifying entities relevant to user questions and providing
targeted responses.
7. Document Categorization: NER helps categorize documents based on
named entities mentioned in the text, enabling content organization and
retrieval.
8. Biomedical and Clinical Text Mining: In the medical domain, NER is
used to identify and extract biomedical entities, such as genes, proteins,
diseases, and drugs, from research papers and clinical notes.

Real‑World Applications of Entity Linking:

1. Knowledge Base Population: Entity linking is crucial for knowledge base

population, where mentions of entities in unstructured text are linked to their
corresponding entries in a structured knowledge base, such as Wikipedia or
DBpedia.
2. Cross‑Document Coreference: Entity linking is used to resolve cross‑doc‑
ument coreference, where mentions of the same entity in different docu‑
ments are linked together.
3. Question Answering Systems: In question answering systems, entity link‑
ing is employed to link named entities mentioned in the question to their
corresponding knowledge base entries, which improves the accuracy of the
system’s responses.
4. News Analysis and Event Tracking: Entity linking aids in analyzing
news articles and tracking events by linking mentions of entities to their
262 Data-Centric Artificial Intelligence for Multidisciplinary Applications

representations in a knowledge base. It helps in understanding the relation‑

ships between entities and events.
5. Semantic Web and Linked Data: Entity linking plays a vital role in the
Semantic Web and Linked Data initiatives, where unstructured textual data
is linked to structured data on the web, creating a network of interconnected
information.
6. Information Retrieval: Entity linking helps in improving information
retrieval systems by linking mentions of entities in documents to relevant
knowledge base entries.
7. Data Integration: In data integration tasks, entity linking is used to inte‑
grate data from different sources by linking mentions of entities to a unified
representation.
8. Machine Reading Comprehension: Entity linking is utilized in machine
reading comprehension tasks to link entities in a passage to their corre‑
sponding knowledge base entries.

Overall, NER and Entity Linking are powerful NLP techniques with various
real‑world applications across diverse domains. They contribute to improving infor‑
mation retrieval, knowledge extraction, and understanding in the era of big data and
unstructured text.

16.7 DEEP LEARNING FOR TEXT MINING

16.7.1 Introduction to Deep Learning for NLP
Deep Learning is a form of machine learning that employs multiple‑layer neural net‑
works to learn and represent complicated patterns in data. Deep learning has dem‑
onstrated exceptional effectiveness in a variety of text mining applications, including
NLP, sentiment analysis, text classification, machine translation, text generation, and
others.2 Deep learning is powerful because it can automatically build hierarchical
representations of text data, incorporating both local and global semantic links. In
Figure 16.7, the operational mechanism of employing Deep Learning for the purpose
of Text Mining is visually elucidated.6
Deep Learning Architectures for Text Mining:

1. Recurrent Neural Networks (RNNs): RNNs are widely used for sequen‑
tial data processing, making them suitable for handling text sequences.
They have a feedback mechanism that allows them to maintain a hidden
state and consider the context from previous words while processing each
word in a text. However, RNNs suffer from the vanishing gradient problem,
limiting their ability to capture long‑range dependencies in texts.
2. Long Short‑Term Memory (LSTM): LSTMs are a type of RNN that
addresses the vanishing gradient problem. They introduce memory cells
that allow information to be stored and retrieved over long periods, making
them better at handling long sequences and capturing long‑term dependen‑
cies in text.5
Machine‑Learning Techniques for Effective Text Mining 263

FIGURE 16.7 Deep learning for text mining.

264 Data-Centric Artificial Intelligence for Multidisciplinary Applications

3. Gated Recurrent Units (GRUs): GRUs are another variant of RNNs

designed to address the vanishing gradient problem. They have fewer
parameters than LSTMs and are computationally more efficient, making
them popular choices for text‑processing tasks.
4. Convolutional Neural Networks (CNNs): Originally developed for image
processing, CNNs have been successfully applied to text mining tasks,
particularly for text classification. They use convolutional layers to cap‑
ture local patterns in text, such as n‑grams, and pooling layers to aggregate
information globally.
5. Transformer‑based Models: Transformers have revolutionized NLP and
text mining tasks. They use self‑attention mechanisms to process entire
sequences of words simultaneously, capturing long‑range dependencies
effectively. The Transformer architecture is the basis for state‑of‑the‑art
models like BERT, GPT‑3, and RoBERT.

Benefits of Deep Learning for Text Mining:

1. Automatic Feature Learning: Deep learning models can automatically

learn useful representations and features from raw text data, reducing the
need for manual feature engineering.
2. Semantic Understanding: Deep learning models excel at capturing com‑
plex semantic relationships between words and sentences, enabling them to
understand the context and meaning in texts.
3. Transfer Learning: Pretrained language models like BERT and GPT‑3
can be fine‑tuned on specific text mining tasks, leveraging their knowledge
from a large corpus to improve performance on domain‑specific data.
4. Handling Sequence Data: Deep learning models like RNNs and
Transformers are well‑suited for processing sequential data like text, where
the order of words is essential.

Challenges:

1. Data Requirements: Deep learning models require large amounts of

labeled data for training, which may be challenging to obtain for certain
specialized domains.2
2. Computational Resources: Training deep learning models can be com‑
putationally intensive and may require significant computational resources
and time.
3. Interpretability: Deep learning models are often complex and lack inter‑
pretability, making it difficult to understand the exact reasoning behind
their predictions.

Despite these challenges, deep learning has demonstrated remarkable advancements

in text mining and continues to be a driving force in NLP research and applica‑
tions. With ongoing research and development, deep learning models are likely to
play an increasingly vital role in understanding and processing text data for various
real‑world applications.
Machine‑Learning Techniques for Effective Text Mining 265

16.7.2 Case Studies Showcasing the Effectiveness of

Deep Learning in Text Mining Tasks
16.7.2.1 Case Study 1: Text Classification using CNN and LSTM
Objective: Classify news articles into predefined topics using deep learning models.
Data: A dataset of news articles with corresponding topic labels (e.g., Sports,
Politics, Technology, etc.).
Approach:

1. Data Preprocessing: Clean the text data by removing stopwords and spe‑
cial characters, and perform tokenization.
2. Word Embeddings: Convert the preprocessed text into word embeddings
using techniques like Word2Vec or GloVe, representing each word as a
dense numerical vector.
3. Model Architecture: Utilize a combination of CNN and LSTM layers to
capture both local and global features in the text data. The CNN layers iden‑
tify local patterns (e.g., n‑grams), while LSTM layers process the sequential
information.
4. Model Training: Train the deep learning model on the labeled dataset
using cross‑entropy loss and backpropagation. Fine‑tune the model on the
training data to optimize the classification performance.
5. Evaluation: Evaluate the model on a separate test dataset using metrics like
accuracy, precision, recall, and F1 score.

Results: The deep learning model achieves high accuracy in classifying news articles
into their respective topics. The combination of CNN and LSTM allows the model
to effectively capture relevant features and patterns in the text, leading to improved
performance compared to traditional machine‑learning methods.

16.7.2.2 Case Study 2: Sentiment Analysis with BERT

Objective: Perform sentiment analysis on customer reviews to determine the senti‑
ment (positive, negative, or neutral) expressed in the text.
Data: A dataset of customer reviews with corresponding sentiment labels.
Approach:

1. Data Preprocessing: Clean the text data by removing noise, special char‑
acters, and stopwords.
2. BERT Embeddings: Use pretrained BERT (Bidirectional Encoder
Representations from Transformers) to convert the text data into contextu‑
alized word embeddings.
3. Fine‑tuning BERT: Fine‑tune the pretrained BERT model on the senti‑
ment analysis task using the labeled dataset. Update the model’s weights to
adapt to the specific sentiment classification task.
4. Model Training: Train the fine‑tuned BERT model on the labeled dataset,
using categorical cross‑entropy loss and gradient descent optimization.
5. Evaluation: Evaluate the BERT‑based sentiment analysis model on a sepa‑
rate test dataset using metrics like accuracy, precision, recall, and F1 score.
266 Data-Centric Artificial Intelligence for Multidisciplinary Applications

Results: The fine‑tuned BERT model demonstrates superior performance compared

to traditional sentiment analysis models. BERT’s contextual embeddings capture
nuances in the text, leading to more accurate sentiment predictions and better gener‑
alization across different domains.
Conclusion: These case studies demonstrate the effectiveness of deep learning
in text‑mining tasks. By leveraging deep learning models like CNN, LSTM, and
BERT, text classification and sentiment analysis tasks can achieve higher accuracy
and better generalization compared to traditional machine‑learning methods. Deep
learning has revolutionized text mining and NLP, allowing for a more sophisticated
and context‑aware understanding of textual data, enabling a wide range of applica‑
tions in various domains.

16.8 CHALLENGES AND ETHICAL CONSIDERATIONS

16.8.1 Challenges in Deep Learning for Text Mining
1. Data Quality and Quantity: Deep learning models require large amounts
of labeled training data to achieve optimal performance. Obtaining
high‑quality labeled data can be challenging, especially for specialized
domains where expert annotations are necessary.2
2. Overfitting: Deep learning models are prone to overfitting, where they mem‑
orize the training data and perform poorly on unseen data. Regularization
techniques and data augmentation are used to address this issue.
3. Interpretability: Deep learning models are often considered black boxes,
making it challenging to interpret their decisions and understand the rea‑
soning behind their predictions. This lack of interpretability can be a sig‑
nificant concern in critical applications.
4. Computation and Resources: Training deep learning models, especially
large‑scale ones like Transformers, can be computationally intensive and
require significant computational resources and time.
5. Hyperparameter Tuning: Deep learning models have many hyperparam‑
eters, and finding the optimal combination can be time‑consuming and
requires expertise.
6. Domain Adaptation: Deep learning models may not perform well when
applied to different domains or when the distribution of the test data differs
significantly from the training data.

16.8.2 Ethical Considerations in Deep Learning for Text Mining

1. Bias and Fairness: Deep learning models can inherit biases present in the
training data, leading to biased predictions. This can perpetuate societal
biases and lead to unfair treatment or discrimination.1
2. Privacy: Deep learning models trained on sensitive text data, such as per‑
sonal conversations or medical records, can raise privacy concerns. Proper
data anonymization and data access controls are essential to protect user
privacy.
Machine‑Learning Techniques for Effective Text Mining 267

3. Misinformation and Fake News: Deep learning models can be used to

generate fake text, leading to the spread of misinformation and fake news.
Ethical considerations involve ensuring the responsible use of text genera‑
tion capabilities and implementing fact‑checking mechanisms.
4. Regulatory Compliance: Deep learning models used in text mining must
adhere to legal and regulatory requirements regarding data privacy, secu‑
rity, and fair use.
5. Data Ownership and Consent: The use of text data for training deep
learning models should comply with data ownership rights, and consent
should be obtained from users whose data is used.
6. Safety and Security: In certain applications, like chatbots or virtual assis‑
tants, deep learning models should be designed to avoid harmful behavior
or providing harmful information.5
7. Transparency and Accountability: Organizations using deep learning
models in text mining should be transparent about their model’s capabili‑
ties, limitations, and potential biases. They should also be accountable for
the outcomes of their applications.
8. Social Impact: Deep learning models deployed in text mining can have sig‑
nificant societal impacts, both positive and negative. Ethical considerations
should ensure that these impacts are carefully assessed and managed.

Addressing these challenges and ethical considerations requires a comprehensive and

multidisciplinary approach, involving collaboration between data scientists, domain
experts, ethicists, and policymakers. Responsible development and deployment of
deep learning models for text mining are crucial to ensure the technology benefits
society while minimizing potential risks and harms.

16.9 EMERGING TRENDS AND FUTURE DIRECTIONS

IN MACHINE LEARNING FOR TEXT MINING
Machine learning for text mining has seen rapid advancements in recent years, driven
by the growing availability of large text datasets, improved deep learning architec‑
tures, and innovative research.1 Several emerging trends and future directions are
shaping the field of text mining and its applications:

1. Transfer Learning and Pretrained Models: Transfer learning has become

a game‑changer in text mining. Pretrained language models like BERT,
GPT‑3, and RoBERTa have shown impressive results by learning rich con‑
textual representations from vast amounts of data. Fine‑tuning these models
on domain‑specific tasks has become a popular approach to leverage their
capabilities for downstream applications.
2. Multimodal Learning: The integration of text with other modalities,
such as images, audio, or videos, is gaining traction. Multimodal learning
enables machines to understand and analyze the context of text in a more
comprehensive manner, opening up new possibilities for tasks like visual
question answering, speech recognition with text comprehension, and more.
268 Data-Centric Artificial Intelligence for Multidisciplinary Applications

3. Explainable AI for Text Mining: The interpretability of deep learning

models is a significant concern. Advancements in explainable AI techniques
aim to make deep learning models more interpretable, allowing users to
understand the reasoning behind model predictions. This is particularly cru‑
cial in critical applications and compliance with regulatory requirements.
4. Low‑resource and Few‑shot Learning: Research is focusing on develop‑
ing machine‑learning models that can perform well with limited labeled
data, known as low‑resource learning. Few‑shot learning techniques aim
to train models with minimal labeled data, making them more adaptable to
new tasks with limited samples.
5. Domain Adaptation and Transfer Learning: Improving model perfor‑
mance on target domains with limited labeled data is an active area of
research. Techniques like domain adaptation and transfer learning aim to
bridge the gap between different domains and leverage knowledge learned
from one domain to benefit another.
6. Bias Mitigation and Fairness: Addressing bias in text mining is critical
to ensure fair and unbiased decision‑making. Researchers are working on
techniques to mitigate bias and promote fairness in text‑mining applications,
particularly when making predictions that may impact individuals or groups.
7. Multilingual and Cross‑lingual Text Mining: With the increasing global‑
ization of information, there is a growing demand for text mining techniques
that can handle multiple languages and cross‑lingual data. Multilingual pre‑
trained models and cross‑lingual transfer learning are becoming essential
components of text mining pipelines.
8. Continual Learning and Lifelong Learning: Traditional machine‑learn‑
ing models often forget previously learned knowledge when trained on new
data. Continual learning and lifelong learning aim to develop models that
can learn from new data while retaining knowledge from previous tasks,
enabling more efficient and adaptable systems.
9. Real‑time and Online Learning: Text mining applications that require
real‑time processing, such as social media monitoring or chatbots, demand
models that can learn and update incrementally. Online learning approaches
are being explored to enable models to adapt to changing data streams.2
10. Unsupervised and Self‑supervised Learning: Reducing the dependency
on labeled data is an ongoing challenge. Unsupervised and self‑supervised
learning methods seek to leverage the abundance of unlabeled data to improve
model performance, making text mining more scalable and accessible.
11. Privacy‑Preserving Techniques: Protecting user privacy and sensitive
information in text mining is gaining significant attention. Research is
focusing on privacy‑preserving techniques like differential privacy and fed‑
erated learning to safeguard individual data while still enabling meaningful
insights.
12. Reinforcement Learning for Text Mining: Reinforcement learning, which
involves training models to interact with an environment and learn from feed‑
back, is finding applications in text generation and dialogue systems. It holds
promise for developing more natural and interactive text‑based interfaces.
Machine‑Learning Techniques for Effective Text Mining 269

The future of machine learning for text mining is exciting, with continuous advance‑
ments unlocking new possibilities and applications. As the field progresses, ethical
considerations, responsible AI development, and societal impacts will remain essen‑
tial components of shaping the direction of text mining research and its transforma‑
tive applications.5

16.10 CONCLUSION
In this chapter, we have explored the dynamic and transformative field of “Machine
Learning Techniques for Effective Text Mining.” Text mining, powered by machine
learning and deep learning, has emerged as a powerful tool to extract valuable
insights from unstructured text data, enabling us to make informed decisions and
understand human language in novel ways.
We began by delving into the definition and scope of text mining, understanding
its importance in processing vast amounts of textual data and extracting meaningful
information from it. We discussed how text mining is employed in various real‑world
applications, including sentiment analysis, NER, text classification, clustering, and
more.
Supervised learning algorithms, such as Naive Bayes, SVMs, Decision Trees,
and Neural Networks, showcased their effectiveness in text classification tasks. We
explored the process of preprocessing and text representation, laying the foundation
for building accurate and robust text classification models.
The chapter then shifted its focus to unsupervised learning techniques, where
algorithms like K‑means, Hierarchical Clustering, and DBSCAN have been instru‑
mental in discovering patterns and structures in unlabeled text data through text
clustering.
Deep learning stole the spotlight as we delved into its application in text mining.
RNNs, LSTM, and the revolutionary Transformers have redefined NLP and pushed
the boundaries of text understanding and generation.
Real‑world case studies illustrated the effectiveness of machine‑learning tech‑
niques in solving practical text mining challenges, showcasing their significance in
diverse domains, including social media monitoring, customer feedback analysis,
sentiment analysis in news articles, and more.
However, we also acknowledged the challenges and ethical considerations in text
mining. Data quality, interpretability, bias, and privacy emerged as key concerns that
demand careful attention and responsible AI development.
As we look to the future, emerging trends in transfer learning, multimodal learn‑
ing, and low‑resource learning promise to shape the landscape of text mining, open‑
ing up new possibilities and applications. Ethical considerations, transparency, and
accountability will remain paramount in driving the responsible advancement of text
mining technologies.
In conclusion, the fusion of machine learning and text mining has paved the
way for unprecedented opportunities in understanding and utilizing textual data.
As this field continues to evolve, researchers, practitioners, and policymakers must
collaborate to ensure the ethical and responsible development of text mining tech‑
niques, harnessing the potential of machine learning to empower us with insights and
270 Data-Centric Artificial Intelligence for Multidisciplinary Applications

knowledge from the vast realm of human language. The journey of text mining is one
of continuous exploration, discovery, and innovation, and its impact on society is set
to grow in profound ways in the years to come.

REFERENCES
[1] Cohen, Aaron M., and William R. Hersh. “A survey of current work in biomedical text
mining.” Briefings in Bioinformatics 6, no. 1 (2005): 57–71.
[2] Doğan, Emre, K. Buket, and Ahmet Müngen. “Generation of original text with text min‑
ing and deep learning methods for Turkish and other languages.” In: 2018 International
Conference on Artificial Intelligence and Data Processing (IDAP), pp. 1–9. IEEE,
2018.
[3] Lewis, David D., Yiming Yang, Tony Russell‑Rose, and Fan Li. “Rcv1: A new bench‑
mark collection for text categorization research.” Journal of Machine Learning
Research 5, no. (2004): 361–397.
[4] Albert, Noel, and Matthew Thomson. “A synthesis of the consumer‑brand relationship
domain: using text mining to track research streams, describe their emotional associa‑
tions, and identify future research priorities.” Journal of the Association for Consumer
Research 3, no. 2 (2018): 130–146.
[5] Jin, Gang. “Application optimization of NLP system under deep learning technology in
text semantics and text classification.” In: 2022 International Conference on Education,
Network and Information Technology (ICENIT), pp. 279–283. IEEE, 2022.
[6] Zhao, Bei, and Wei Gao. “Machine learning based text classification technology.”
In: 2022 IEEE 2nd International Conference on Mobile Networks and Wireless
Communications (ICMNWC), pp. 1–5. IEEE, 2022.
 17 Emails Classification
and Anomaly Detection
using Natural Language
Processing
Tanvi Mehta, Renu Kachhoria,
Swati Jaiswal, Sunil Kale, Rajeswari Kannan,
and Rupali Atul Mahajan

17.1 INTRODUCTION
The Enron Corporation, an American energy business with headquarters in Houston,
Texas, and amongst the five largest accountancy and audit firms in the world, filed
for bankruptcy as a result of the Enron controversy, which became public in 2001 [1].
Enron was noted as the worst audit failure at the time in addition to being the big‑
gest bankruptcy reorganization in American history. The majority of its clients had
left, and the business had eventually stopped functioning [2]. Despite suffering
billion‑dollar losses in pensions and asset prices, Enron’s employees and stockhold‑
ers only won little compensation through litigation.
Customer satisfaction (CS) assessment is now a key metric for assessing the suc‑
cess of businesses in the market [3]. Client satisfaction has become the top priority
for all business kinds. Analyzing client reviews and comments for a product or ser‑
vice is one technique to gauge CS [4]. Domo claims that we produce more than 2.5
quintillion bytes of data every day, with this data generated from a variety of sources,
including social media, emails, Amazon, YouTube, and Netflix.
The earliest method of business communication is said to be email. It is inevitable
in a workflow scenario that involves both internal and external actors and is defined
by both [5]. The email bodies inherently include the characteristics of the organi‑
zational process viewpoint. The organizational model, however, has not been fully
taken into account in numerous publications that concentrated on email analysis for
process model mining [6]. So, this research proposes ways like Anomaly Detection,
Social Network Analysis, Email Classification, and Word Cloud, that are employed
using Machine Learning, and Natural Language Processing to evaluate the informa‑
tion and assist toward corporate development.
Anomaly detection, also known as outlier detection, is a technique used in data
mining to locate unusual things that happen or observations that stand out from the
bulk of the data in a way that raises questions [7]. Generally, anomalous objects will

DOI: 10.1201/9781003461500-20 271

272 Data-Centric Artificial Intelligence for Multidisciplinary Applications

point to some sort of issue, like financial fraud, a structural flaw, a health issue, or
syntax errors in a text. For instance, sudden bursts of activity rather than infrequent
items are frequently intriguing objects in the context of abuse and network intrusion
detection [8]. Isolation Forest is used to create a model of typical behavior and iden‑
tify unusual conduct before the public controversy.
Finding the communities inside a virtual community is crucial because it enables
the identification of members who have common interests and behavior prediction [9].
Social network analysis is a method for analyzing social systems utilizing networks
and graph theory [10]. By changing the visual representation of a network’s nodes and
edges to reflect certain properties of interest, these visualizations offer a method for
qualitatively evaluating networks. Additionally, emails in the collection are categorized
into documentation, transactions, attorney, etc. This is accomplished by first employing
a bag of words, then SVM, Naive Bayes, and RNN approaches. Furthermore, a word
cloud is created for visualization and comparative analysis is done.
A survey of the preceding scholarly papers is provided in Section 17.2. A compre‑
hensive outline of the methods employed in the aforementioned investigations is pro‑
vided in Section 17.3. An explanation of the approaches used is given in Section 17.4.
Section 17.5 offers an in‑depth description of how the suggested tasks are executed.
In Section 17.6, the conclusions and observations of the research are in‑depth ana‑
lyzed. Section 17.7 addresses the outcome and prospective applications.

17.2 LITERATURE REVIEW

Hyenkyun Woo., in 2021 [11], introduced a matrix factorization technique that uses
low dimensional estimates of the source data to detect abnormalities in a natural way.
Block Coordinate Descent is the foundation of the iterative method of Text Outliers
using Nonnegative Matrix Factorization. The term‑document matrix is defined in
blocks so that the function may be solved. They always changed one matrix unit to
the best possible state using the most recent updates to the values of the other blocks.
The method resulted in an accuracy of 93.40%.
Vipin Kumar et al., in 2017 [12], intended to offer a detailed and organized sum‑
mary of the anomaly detection studies. They divided current approaches into many
groups according to the fundamental philosophy behind each technique. They identi‑
fied important presumptions for each category that the approaches rely on to distin‑
guish between typical and abnormal behavior. These presumptions can be utilized as
guides to determine if a particular approach is effective in a given area when applied
to it.
Philip Branch et al., in 2020 [13], employed a triangle model with three vertices to
represent the problem (large demography), methods/algorithms (outliers), and tools
(big data software solutions), and described the situation of feature extraction in large
features of data. Moreover, recent methodologies and applications necessary for the
enhancement of outlier detection were examined along with the shortcomings of
classic approaches and current tactics of high‑dimensional data.
Nisheeth Srivastava et al., in 2022 [14], showed how knowledge of the linguistic
contents of textual data enables the identification of word and subject occurrences
that are statistically uncommon but contextually reasonable for the model being
Emails Classification and Anomaly Detection 273

monitored, and thus, not abnormal. It was accomplished by bootstrapping a unique

textual clustering technique for outlier identification that utilizes Linear discriminant
analysis (LDA) on specific datasets to a frame of reference method that works on
independent corpora of generic text connections (WordNet). The algorithm gave an
accuracy of 88.64%.
Anuradha Pillai et al., in 2017 [15], reviewed several natural language and text
mining methodologies to extract a social network from text data. These tech‑
niques include Automatic Summarization, Chunking, Named Entity Relation,
Disambiguation, and Fact Extraction. Furthermore, they analyzed the challenges of
implementation corresponding to every technique. They concluded that linguistics
techniques will make searches more user‑friendly, and mining text will tap into its
intelligence.
Sotiris Kotsiantis et al., in 2015 [16], classified the text by utilizing many machine
learning techniques. They primarily focused on the process of classifying the text
which includes Text data, tokenization, Stemming, Representing Vectors, Feature
Selection and Transformation, and using algorithms. By first examining the quali‑
ties of training corpora and then providing an approach for creating learning text
semi‑automatically, the authors attempted to provide a method for creating high‑qual‑
ity training corpora for better classification performance.
Kelly Wilkinson et al., in 2021 [17], investigated the patterns of text in the graphi‑
cal form of a word cloud representing the knowledge of students. This approach was
implemented to let the instructors evaluate if the students have grasped the concepts
or made any mistakes. They analyzed that it can also be utilized to give academic
feedback to students.
Sidra Abbas et al., in 2022 [18], categorized the text using deep learning tech‑
niques. They employed artificial neural network, long short‑term memory (LSTM),
and gated recurrent unit for classification. LSTM model outperformed with the high‑
est accuracy of 92% on the Titanic textual data.
Tiantian Zhang et al., in 2021 [19], implemented an isolation‑based technique
called Isolation Forest Algorithm to detect the anomaly points. An elevated algo‑
rithm, Cluster‑Based Isolation Forest based on this approach translates localized out‑
liers before clustering into global outliers of nearby collections and then determines
the outlier score for every cluster’s data point. This research increased the accuracy
by 6%.

17.3 TECHNIQUES
The proposed techniques for analyzing the emails of the Enron organization are
briefly explained in this section.

17.3.1 Isolation Forest
Employing a tree‑like framework in accordance with traits selected arbitrarily, an
isolation forest is applied to examine randomly subsampled data [20]. The fragments
that crossed farther into the tree required more cuts to separate, thus making them
less probable to be anomalies [21]. Similar to the last example, data that end up on
274 Data-Centric Artificial Intelligence for Multidisciplinary Applications

shorter branches tend to be anomalies since the tree found it simpler to distinguish
them from others.
E ( h ( x ))
−
S ( x, n) = 2 c( n )
(17.1)

where S (x, n) = score of an anomaly; E (h(x)) = observations’ path length x; c(n) = unsuc‑
cessful search path length that is average; n = count of nodes that are external.

17.3.2 Support Vector Machine (SVM)

SVM can be applied to address categorization challenges. In this supervised
approach, data are separated using the hyperplane, which acts as the ideal selection
boundary [22]. It identifies the extreme points, termed as support vectors, to help
build the hyperplane. It also has a positive hyperplane that crosses a minimum of one
of the closest positive indicators, as well as a negative hyperplane that traverses either
of them or more of the nearby negative points [23]. The ideal hyperplane is the one
where, as demonstrated, the margin – the space between the ideal hyperplane, and
the ideal hyperplane is at its greatest.

17.3.3 Naïve Bayes (NB)

NB, a guided method, is used to solve classification problems and is built on the
Bayes Theorem. It is mostly used to classify text, including large training datasets
[24]. It assists in creating efficient and precise machine learning models and produc‑
ing prompt forecasts.

17.3.4 Recurrent Neural Network (RNN)

A sort of neural network called an RNN uses the output of a previous stage as an
input for the current stage [25]. Inputs and outputs in traditional neural networks are
distinct from each other, but when it comes to predicting the subsequent word in a
phrase, it is vital to remember the prior phrases. As a result, RNN was developed,
and it employed a Hidden Layer to address this problem [26]. The Hidden state of
RNNs, which saves specific information about a sequence, is its central and most
significant feature.

17.3.5 Natural Language Toolkit (NLTK)

One of the most popular platforms for creating Python applications that use human
language data is called NLTK [27]. In addition to a selection of text‑examining librar‑
ies for categorizing, encoding, rooting, labeling, parsing, and linguistic thinking for
industrial‑strength NLP libraries and an active discussion forum, it offers simple
designs for accessing more than 50 corpus and semantic assets including WordNet.
Emails Classification and Anomaly Detection 275

17.3.6 Gensim
Gensim is an unrestricted Python toolkit that aims to represent documents as seman‑
tic vectors as quickly and painlessly as possible for humans and computers [28].
Gensim uses unsupervised machine learning methods to analyze unstructured,
uncooked digital messages.

17.3.7 Bag of Words (BoW)

When text is being modeled using deep learning and machine learning techniques,
BoW is a method of expressing text data [29]. This model counts the number of
times each word appears to convert any text into fixed‑length vectors. This approach
is referred to as the vectorization process. By extracting text characteristics, this
straightforward method is used to address issues with language modeling and docu‑
ment categorization [30]. It provides a great deal of customization options for certain
text data.

17.4 PROPOSED FRAMEWORK

The usual implementation approach for email text analysis is outlined in this section.
Figure 17.1 depicts an illustration of the research methodology. The procedure is
briefly covered below:

17.4.1 Email Dataset
The Enron email dataset consists of around 0.5 million emails between the employ‑
ees and managers of the company. This data is needed to be cleaned and processed
for feature extraction and selection.

17.4.2 Text and Visual Feature Extraction

For feature extraction various natural language processing techniques like
Tokenization and Lemmatization are used. Bag of Words is employed on the text
data to extract vital features.

17.4.3 Feature Selection

Features are selected on the basis of POS Tagging, stopwords removal, etc. The gen‑
ism library of natural language processing is used for this purpose.

17.4.4 Model Formation
Several techniques are used for examining the data of emails, including anomaly
detection, social network analysis, email classification, and word cloud formation.
276 Data-Centric Artificial Intelligence for Multidisciplinary Applications

FIGURE 17.1 Framework for the pipeline.

17.5 PROPOSED METHODOLOGY

17.5.1 Problem Statement
One of the biggest business collapses in history was caused by the Enron scandal and
collapse. Among the major energy firms in America before 2000 was Enron. After
being exposed for fraud, it then plummeted into insolvency in less than a year. There
are 150 former Enron workers, most of whom are top executives, and around 500,000
emails between them. Its value is increased by the fact that it is the only sizable
public database of authentic emails. In reality, data scientists have been studying and
researching with this dataset for years. This experimentation puts forward tasks like
Anomaly Detection, Social Network Analysis, Email Classification, and Word Cloud
generation that are to be performed on the dataset for visualization.

17.5.2 Anomaly Detection
Finding patterns in the data that do not match the anticipated (normal) behavior is
known as anomaly detection. Novelties, noise, outliers, exceptions, and deviations
Emails Classification and Anomaly Detection 277

FIGURE 17.2 Algorithmic flow of isolation forest.

are other terms for anomalies. The algorithmic flow of the Isolation Forest technique
is demonstrated in Figure 17.2.

17.5.2.1 Importing Libraries

Various libraries like numpy, pandas, matplotlib, seaborn, email, tqdm were imported
at the initial stage.

17.5.2.2 Data Preprocessing and Creating Data Frame

This step involves all needed transforms to interpret the data and then use it for
future tasks. Converting data to table format for running various queries and pro‑
cesses is required.

17.5.2.3 Adding Content Length as a Column

This is a crucial column for analysis since it shows how long each email is and will
be highly useful for highlighting data points that deviate from the overall trends.
Typically, emails that are not responses are substantially greater in size than those
that are.

17.5.2.4 Adding a Column for Whether an Email Is a Reply or Not

Later, this column will serve as a field for finding anomalies. It also shows the num‑
ber of brief exchanges occurring as a feature in the dataset, which will be shown
later. The column also indicates that, as will be shown later, letters in this column
tend to be shorter than usual.
278 Data-Centric Artificial Intelligence for Multidisciplinary Applications

17.5.2.5 Generating Time‑Series Data

In order to achieve this, the data must be grouped by day and hour. After that, time
series analysis and anomaly detection may be done using this data. For different
fields, an aggregate like a mean or count is often taken, and the results are displayed
as and when they are applied to all emails received during that specific hour.
The number of mails shoots up very fast in a very specific time frame which
indicates anomalous behavior in that period of 2–3 days, as shown in Figure 17.3.
Similar anomalous hours are also seen about a week after that. This can also be a
strong indicator of fraudulent activities like the deletion of large amounts of records
and other instructions being passed and expecting immediate reporting on the same.
Sharp anomalous surges in the number of emails to be preceded by surges in the
length of emails thus it is also a very important indicator of fraudulent activity, as
shown in Figure 17.4.
The number of replies in most conditions is very low or extremely high depending
on the hour being considered, which can be viewed in Figure 17.5. These points of
multiple replies on the same topic can be of key interest in some cases.
Figure 17.6, shows that in most cases anomalies are in one of the three dimensions
and rarely in all of them together. Some points are slightly anomalous in terms of
length and replies but do not deviate significantly.

17.5.2.6 Performing Anomaly Detection Using Isolation Forests:

Initially, IsolationForest library is imported from sklearn. It creates trees using data
and arbitrary splits. Anomalies will be distinct from these random separations since
they will be removed from the data.

17.5.2.7 Considering Data Points for Analysis

Using 0.1 as a threshold and few data points are considered, and anomalies are
printed.

FIGURE 17.3 Drastic change within 2–3 days.

Emails Classification and Anomaly Detection 279

FIGURE 17.4 Sharp surges of emails.

FIGURE 17.5 Conditions on an hourly basis.

17.5.2.8 Using Local Outlier Factor

First, LocalOutlierFactor library is Imported using the sklearn library. Novelty helps
in detecting new features using k‑nearest neighbors, and n‑neighbors help in deciding
the density for outliers. Data seems to have a very high density in two major regions,
as shown in Figure 17.7. So, this technique doesn’t seem to work very well, and thus
Isolation forests will be used.

17.5.2.9 Creating a Count of Mail Sent by Every Source and Receiver

Most of the mails have been sent by the top 200 employees, as seen in Figure 17.8,
and thus it would be more relevant to consider only their data in general so that exter‑
nal factors don’t influence them significantly.
280 Data-Centric Artificial Intelligence for Multidisciplinary Applications

FIGURE 17.6 Anomalous behavior in a 3D graph.

FIGURE 17.7 Anomalies depicted using local outlier factor.

Figure 17.9 shows that top the 500 recipients have actually received most of the
emails in the organization.

17.5.2.10 Creating a Count of Every Pair of Source and Destination

Creating a new feature for every pair involved in a mail. This will help in identifying
new novel communication patterns.

17.5.2.11 Finding Unique External Emails

Unique external senders and recipients have been noted as they are vents for the
organizational information to escape.
Emails Classification and Anomaly Detection 281

FIGURE 17.8 Senders.

FIGURE 17.9 Receivers.

17.5.2.12 Creating Data and Anomaly Detection

Creating a numpy array that has all features to consider for final anomaly detection,
This involves printing all columns being used as features and adding number of
unique sources and several emails too per hour. Then, the final anomaly detection
system using Isolation Forests was detected. When the number of emails is factored
into the features; most features end up becoming white noise and thus are redundant.
These features can then be removed. Hence, the anomalies are detected.

17.5.3 Social Network Analysis

Networked structures are described in terms of nodes and the ties, edges, or links
connecting them. Social media networks, the dissemination of memes, the flow of
282 Data-Centric Artificial Intelligence for Multidisciplinary Applications

information, friend and acquaintance networks, and commercial networks are a

few examples of social structures that are frequently represented by social network
analysis.

17.5.3.1 Importing Libraries

Various libraries like numpy, pandas, seaborn, matplotlib are imported that are used
for data preprocessing and visualization.

17.5.3.2 Data Extraction

Data is imported into the Pandas Dataframe which is then used for further computa‑
tions and analysis.

17.5.3.3 Functions to Extract Information from Mail

Libraries like tqdm are imported to extract the information. It returns the series with
text sliced from a list split from each message. After this, the mails that are not in the
proper structure are cleaned or removed.

17.5.3.4 Analysing the Social Network

To examine the social network libraries like network, and nxviz are imported.
The Arc plot can be visualized in Figure 17.10. It can be seen that the nodes or
employees are at the bottom and the dominance of the one node is on the right side.
Since the dataset contains about 5 lakhs of emails, the emails from the 150 Enron
executives contained in the full dataset are considered.
The Circos plot is shown in Figure 17.11. It is clearer and easier to comprehend
this plot. Finally, the social network is drawn as part of the result.

17.5.4 Email Classification and Word Cloud

Various techniques are used to classify emails based on different topics. Text prepro‑
cessing is done using NLP, and for classification, SVM, NB, and RNN models are
employed.

FIGURE 17.10 Arc plot.

Emails Classification and Anomaly Detection 283

FIGURE 17.11 Circos plot.

17.5.4.1 Importing Libraries

Many libraries such as numpy, pandas, matplotlib, genism, spacy are imported for
further preprocessing.

17.5.4.2 Tokenization
In this type of text preprocessing concept, first, the text is normalized, converted to
lowercase, punctuations are removed, and finally split into words; these words are
called tokenizers.

17.5.4.3 Stop Words Removal

In this step, the words that are used often across the corpus of documents or sen‑
tences are eliminated. These words are most likely to be “a, an, the, etc.”.

17.5.4.4 Lemmatization and Stemming

Lemmatization and stemming often relate to carrying out procedures correctly using
a lexicon and semantic evaluation of words with the goal of removing only affixation
ends and returning the lemma, or dictionary form, of a word.

17.5.4.5 Sklearn Pipeline Using CountVectorizer() and TfidfTransformer()

The tokenize function is supplied as a parameter when using the CountVectorizer ()
to provide vocabulary and encode new documents with that vocabulary. The opera‑
tor transfers a count vector to a standardized tf illustration. This word weighting
284 Data-Centric Artificial Intelligence for Multidisciplinary Applications

technique is widely used in information retrieval and is very effective in document

categorization.

17.5.4.6 Using Point‑Wise Mutual Information (PMI) for Feature Selection

PMI is a statistical technique frequently used to model the relationship between words.
The following definition of PMI exists between the terms “word1” and “word2”:

 p ( word1& word2 ) 
PMI ( word1, word2 ) = log2   (17.2)
 p ( word1) xp ( word2 ) 

17.5.4.7 Developing Model and Analyzing Evaluation

The model is developed using SVM, NB, and RNN. Their performance is analyzed
using several metrics like accuracy, precision, recall, and F1 score. The formulae are
given as:

TP + TN
Accuracy = (17.3)
TP + TN + FP + FN

TP
Precision = (17.4)
TP + FP

TP
Recall = (17.5)
TP + FN

2TP
F1 Score= (17.6)
2TP + FP + FN

17.5.4.8 Developing a Word Cloud

A text visualization approach called a “word cloud” or “tag cloud” is used to dis‑
play tags or terms from web pages. These keywords are frequently single words that
describe the content of the website from which the word cloud was derived. A Word
Cloud is created by grouping these words collectively.
The font size and color of each word in this word cloud may be changed. A word’s
significance in relation to other words in the cluster is more accurately portrayed by
a word’s larger font size. Depending on the authors’ concept, word clouds may be
constructed in a variety of forms and sizes. The overall number of words is important
when creating a Word Cloud. A Word Cloud with more words is not always better
since it becomes congested and harder to read. A word cloud must always represent
its intended usage appropriately and have semantic significance.

17.5.5 Pseudocode
function: ClassifyEmailText(t): (T, l)
input: t: string ‑ the email text to be distinguished.
Emails Classification and Anomaly Detection 285

output: T = {wo, ..., wn} ‑ an ontology’s set of concepts

that may categorize the present email text.
l: integer – the resemblance between the terms
on a scale.
1 Ω = PreprocessEmailText(t) ‑ generates a list
of words after textual preprocessing
2 l = ∞ ‑ finest resemblance level.
3 T =Ø
4 for each wk Ɛ Ω do
5 Θk = ComputeSimilerLevels(wk,0).
6 for each <wj , lj> Ɛ Θk
7 if lj < l then
8 T = {wj}.
9 l = lj.
10   else if lj = l then
11 T = T ∪{wj}.
12 end‑if
13 end‑for each
14 end‑for each
15 return<T, l>.

17.6 EXPERIMENTAL RESULTS AND ANALYSIS

17.6.1 Results of Anomaly Detection
In Figures 17.12–17.16, it can be seen that most of the outliers are now being detected
and flagged. The number of flagged cases is about 550 hours out of 18,000 hours
which can be analyzed as it is only about 3% of all the cases.

FIGURE 17.12 Content length.

286 Data-Centric Artificial Intelligence for Multidisciplinary Applications

FIGURE 17.13 Reply.

FIGURE 17.14 Outside receiver.

17.6.2 Results of Social Network Analysis

The social network shown in Figure 17.17 has been plotted with different “tension”
on the spring layout to depict different options for visualizing this network.
Emails Classification and Anomaly Detection 287

FIGURE 17.15 Fraction communication.

FIGURE 17.16 Number of emails.

17.6.3 Results of Emails Classification and Word Cloud

Accuracy, recall, precision, and F1 score are three metrics utilized to express the
output of different models. Table 17.1, Figures 17.18 and 17.19 present these outcomes
with efficiency. It demonstrates that RNN achieves the highest accuracy possible of
97.8%.
The word clouds generated using NLP techniques are depicted in Figures 17.20
and 17.21.
288 Data-Centric Artificial Intelligence for Multidisciplinary Applications

FIGURE 17.17 Social network.

TABLE 17.1
Evaluation Metrics
Positive Negative
Models Precision Recall F1 Score Precision Recall F1 Score Accuracy
NB 0.92 0.91 0.89 0.87 0.93 0.91 0.924
SVM 0.85 0.89 0.86 0.98 0.86 0.87 0.873
RNN 0.97 0.97 0.97 0.97 0.97 0.97 0.978

17.7 CONCLUSION
This research suggests different methods for analyzing the email dataset of Enron
company. Techniques like Anomaly Detection, Social Network Analysis, Email
Classification, and Word Cloud generation are used for examining the anoma‑
lies. Enhanced and futuristic technologies like natural language processing and
machine learning are applied to investigate the mails and the social network formed.
Algorithmic techniques like Isolation Forest, NB, SVM, and RNN are used to train
suitable models. The RNN model was the most accurate, reaching 97.8%, based on
the findings. An experiment has been conducted as part of this study to elevate the
model’s analytical outcomes. The goal is to investigate and perform more research
on anomaly detection in line with fraud analysis in various sectors using the hybrid
approach, wherein the dataset may be hosted in a cloud environment like AWS.
Emails Classification and Anomaly Detection 289

FIGURE 17.18 Evaluation metrics for models.

FIGURE 17.19 Accuracy of models.

As a consequence, all the traits and features are analyzed, tested, and the most accu‑
rate findings are obtained. As a result, this will significantly aid in both the growth
and success of organizations as well as the advancement of business.
290 Data-Centric Artificial Intelligence for Multidisciplinary Applications

FIGURE 17.20 Word cloud with white background.

FIGURE 17.21 Word cloud with black background.

REFERENCES
1. N. Shashidhar et al. (2022). Topic modeling in the ENRON dataset. In: Hu B., Xia Y.,
Zhang Y., Zhang L.J. (eds) Big Data‑BigData 2022. BigData 2022. Lecture Notes in
Computer Science, vol. 13730, Springer, Cham. doi:10.1007/978‑3‑031‑23501‑6_4.
2. M. MacDonnell et al. (2022). Exploring social‑emotional learning, school climate, and
social network analysis. Journal of Community Psychology, 51(1), 84–102.
3. R. Benbenishty et al. (2017). A research synthesis of the associations between socioeco‑
nomic background, inequality, school climate, and academic achievement. Review of
Educational Research, 87(2), 425–469.
4. T. Mehta et al. (2022). A comparative study on approaches for text quality predic‑
tion using machine learning and natural language processing. In: 2022 International
Conference on Smart Generation Computing, Communication and Networking
(SMART GENCON), Bangalore, India, 2022, pp. 1–5.
Emails Classification and Anomaly Detection 291

5. V. Sonakshi et al. (2023). An employee feedback model based on sentimental analysis

and word clouds. In: Data Science and Intelligent Computing Techniques, SCRS, India,
pp. 79–92.
6. N. El Gayar et al. (2019). Sentiment analysis using unlabeled email data. In:
2019 International Conference on Computational Intelligence and Knowledge
Economy (ICCIKE), Dubai, United Arab Emirates, pp. 328–333. doi:10.1109/
ICCIKE47802.2019.9004372.
7. T. Mehta et al. (2023). Sentiment analysis of political tweets for israel using machine
learning. In: Machine Learning and Big Data Analytics, ICMLBDA 2022, Proceedings
in Mathematics and Statistics, vol. 401, Springer, New York.
8. M. Kurdi (2022). Identifying and analyzing communities within a social network
using automatic topic labelling: application to the enron dataset. In: 18th International
Conference on Machine Learning and Data Mining, New York, USA.
9. N. Thakur et al. (2020). Evaluating cohesion score with email clustering. In: Proceedings
of First International Conference on Computing, Communications, and Cyber‑Security
(IC4S 2019). Lecture Notes in Networks and Systems, Punjab, India, vol. 121.
10. O. Elisha et al. (2022). MessageNet: message classification using natural language pro‑
cessing and meta‑data. arXiv preprint arXiv:2301.01808.
11. C. Aggarwal et al. (2017). Outlier detection for text data. In: Proceeding of SIAM
International Conference on Data Mining, Houston, Texas, USA.
12. A. Banerjee et al. (2009). Anomaly detection: A survey. ACM Computing Surveys, 41.
doi:10.1145/1541880.1541882.
13. J. Jin et al. (2020). A comprehensive survey of anomaly detection techniques for high
dimensional big data. Journal of Big Data, 7, 42.
14. J. Srivastava et al. (2012). Contextual anomaly detection in text data. Algorithms, 5,
469–489. doi:10.3390/a5040469.
15. D. Juneja et al. (2017). Extracting information from social network using Nlp. The
International Journal of Computational Intelligence Research, 13, 621–630.
16. S. Kotsiantis et al. (2005). Text classification using machine learning techniques.
WSEAS Transactions on Computers, 4, 966–974.
17. K. Wilkinson et al. (2014). Get your head into the clouds: using word clouds for analyz‑
ing qualitative assessment data. TechTrends, 58, 38–44. doi:10.1007/s11528‑014‑0750‑9.
18. S. Abbas et al. (2022). An efficient approach for textual data classification using
deep learning. Frontiers in Computational Neuroscience, 16, 992296. doi:10.3389/
fncom.2022.992296.
19. R. Gao et al. (2019). Uniform mixture model in the case of regression. Journal of
Physics: Conference Series, 1237, 052023.
20. R. S. Prasad (2019). Influence of language‑specific features for author identification
on Indian literature in Marathi. In: Soft Computing and Signal Processing’s of 2nd
ICSCSP 2019. Hyderabad, India.
21. R. Kachhoria et al. (2023). Lie group dee learning technique to identify the precision
errors by map geometry functions in smart manufacturing. The International Journal
of Advanced Manufacturing Technology. 10.1007/s00170‑023‑10834‑2.
22. M. Lee et al. (2019). Use of social network analysis in the development, dissemination,
implementation, and sustainability of health behavior interventions for adults: A sys‑
tematic review. Social Science & Medicine, 220, 81–101.
23. S. Kale et al. (2018). Author identification on imbalanced class dataset of Indian litera‑
ture in Marathi. International Journal of Computer Sciences and Engineering. vol. 6,
pp. 542–547.
24. C. Haythornthwaite et al. (2018). Social network analysis: An approach and technique
for the study of information exchange. Library & Information Science Research, 18(4),
323–342.
292 Data-Centric Artificial Intelligence for Multidisciplinary Applications

25. K. Rao et al. (2017). Text analysis for author identification using machine learning.
Journal of Emerging Technologies and Innovative Research. 4(6), pp.138–141.
26. T. Mehta et al. (2022). YouTube Ad view sentiment analysis using deep learning and
machine learning. International Journal of Computer Applications 184(11), 10–14.
27. S. D. Kale et al. (2017). A systematic review on author identification methods.
International Journal of Rough Sets and Data Analysis (IJRSDA). 4(2), pp. 81–91.
28. K. Chang et al. (2011). Word cloud model for text categorization. IEEE 11th International
Conference on Data Mining, Vancouver, BC, Canada, 2011, pp. 487–496.
29. S. D. Kale et al. (2018). Author identification using sequential minimal optimization
with rule‑based decision tree on Indian literature in Marathi. Procedia Computer
Science. Procedia Computer Science, Elsevier, vol. 35, pp. 1086–1101.
30. F. Chang et al. (2019). Hot topic community discovery on cross social networks. MDPI
Future Internet, 11(3), 60–76.
Index
Adam (Adaptive Moment Estimation) 64 Generative adversarial networks (GANs) 169
aerial surveillance 166 glomerular filtration rate (GFR) 88
Agriculture Data Exchange (ADEx) 173 GoogleNet 134
AlexNet network 103
Alzheimer’s disease (AD) 91 healthy–diseased (HD) 183
anomaly detection 276 hierarchical clustering 248
anonymization 13 Hypertext Markup Language (HTML) 192
artificial neural networks (ANNs) 179
IndCareer.com 191
bag of words (BoW) 218 Indian classical music (ICM) 204
Bidirectional Encoder Representations from International Data Corporation (IDC) 45
Transformers (BERT) 242 Internet of Things (IoT) 160
big medium very big (BMV) 183 isolation forest 273
BLEU (Bilingual Evaluation Understudy) 65
JavaScript 192
Careers360 190 JavaScript Object Notation (JSON) 192
cerebral spinal fluid 134
chronic kidney disease (CKD) 88 K‑means clustering 248
chronic venous insufficiency (CVI) 117
CollegeDekho.com 190 Latent Dirichlet Allocation (LDA) 237
convolution layer 134 Locally Interpretable Model-Agnostic
convolutional neural networks 10 Explanations (LIME) 147
customer satisfaction (CS) 271 long short-term memory (LSTM) 11

data bias and fairness 37 machine learning 7

data‑centric AI 3 max pooling 134
data fusion and integration 36 mild cognitive impairment (MCI) 91
data maintenance 32 multimodal learning 267
data preprocessing 167 music information retrieval (MIR) 202
decision support systems (DSSs) 159
decision trees 242 Naive Bayes 240
deep neural network 133 named entity recognition (NER) 217
Density-Based Spatial Clustering of Applications natural language inference (NLI) 42
with Noise (DBSCAN) 249 natural language processing (NLP) 150
dimensionality reduction 167 natural language toolkit (NLTK) 217
discrete wavelength transform 134 neural network 137
discrete wavelet transform (DWT) 101 Non-negative Matrix Factorization (NMF) 237

Edge AI 17 one-hot coding 218

explainable artificial intelligence (XAI) 142 openpyxl 192

false negative predictive value (FNR) 183 Parsimony 152

false omission rate (FOR) 183 partial least squares regression (PLSR) 179
false positive predictive value (FPR) 183 part‑of‑speech (POS) 245
FDR (false discovery rate) 183 pattern recognition (PR) 202
feature engineering 167 positron emission tomography (PET) 99
flask 192
fuzzy C‑means 134 reconciling data-centric AI 49
recurrent neural networks (RNNs) 262
gated recurrent unit (GRU) 12 reinforcement learning 8
General Data Protection Regulation (GDPR) 146 ReLU (rectified linear unit) 10

293
294 Index

RFID 165 t-distributed Stochastic Neighbor Embedding

ripe unripe (RU) 183 (t-SNE) 250
rule-based machine translation (RBMT) 56 TER (translation edit rate) 66
term frequency-inverse document frequency
SHapley Additive exPlanations (SHAP) 146 (TF-IDF) 218
Shiksha.com 190 tokenization 218
short-term memory (LSTM) 242 total acid content (TAC) 178
SoftMax 134 total soluble solid (TSS) 179
soil sensors 165 transfer learning 267
soluble solids content (SSC) 178 transparency 152
statistical machine translation (SMT) 56
stemming 218 vector space model (VSM) 250
Stochastic Gradient Descent (SGD) 64 visual geometry group (VGG)-19 network 103
support vector machine (SVM) 179