ISSN 2278-3091

Volume 10, No.3, May - June 2021

Syed Hassan Ali Shah et al., International Journal of Advanced Trends in Computer Science and Engineering, 10(3), May - June 2021, 1895 – 1900
International Journal of Advanced Trends in Computer Science and Engineering
Available Online at http://www.warse.org/IJATCSE/static/pdf/file/ijatcse571032021.pdf
https://doi.org/10.30534/ijatcse/2021/571032021

A Review on Anomaly Detection in Time Series

Syed Hassan Ali Shah 1∗, Muhammad Junaid Akbar 2, Usman Ahmed Raza 3
1
Department of Computer Science, Lahore Leads University, Lahore, Pakistan, [email protected]
2
Department of Computer Science, Lahore Leads University, Lahore, Pakistan, [email protected]
3
Department of Computer Science, Lahore Leads University, Lahore, Pakistan, [email protected]

ABSTRACT these fields than they did with simplistic analysis

with less numbers.
Time series is a very common class of data sets.
Among others, it is very simple to obtain time series Although there are relatively few situations in which
data from a variety of various science and finance the standard abnormality seems to be off for time
applications and an anomaly detection technique for series results, this does not inherently indicate that
time series is becoming a very prominent research the abnormalities are not important. Under the tiny
topic nowadays. Anomaly identification covers dataset behind the unusual values, the actual piece of
intrusion detection, detection of theft, mistake data which be hidden. In the field of medicine an
detection, machine health monitoring, network sensor abnormal heart beat is identifiable enough that the
event detection or habitat disturbance. It is also used ECG observation is able to reveal the disorder in
for removing suspicious data from the data set before time. In addition, time-series anomaly detection
production. This review aims to provide a detailed software may be used to map engine state, network
and organized overview of the Anomaly detection violation detection, money-laundering, tracking of
investigation. In this article we will first define what public opinion networks, credit card theft, stock
an anomaly in time series is, and then describe market analysis, unlawful enforcement of tax acts,
quickly some of the methods suggested in the past major inspections of construction sites, possibly any
two or three years for detection of anomaly in time IT device, and several other IT scenarios. And though
series. it could be fascinating to research this phenomena
from a scientific viewpoint, it may make your life
simpler.
Key words : Anomaly Detection, Time Series,
Outlier, Review 2. ANOMALY DETECTION DEFINITION OF
TIME SERIES

1. INTRODUCTION Obtained Time Series are commonly available in

many major databases involving finance, medicine,
The growth of data has increased with growing use of engineering, and social science. It is important to
information technology. The recent study, conducted understand that there are two different features of
in at University of Southern California, has shown time-series data: time-series and other data types.
that after 1980s, the scale of data has risen First, we need to group the records by time and/or
dramatically and doubled per year or even a few day of the year, and before including the records we
months. Both the global data and the regional data must compare the values of the attributes effectively
reached 295EB until 2007, and then they hit about and in a logical order. Secondly, in certain places and
1.8 ZB until 2011, and more than 40 ZB by 2020. Big times, the sequence attribute, the record values, are a
data time has come to university these days. In long flowing sequence. To determine whether time
addition, it offers solutions to figure out how to series are univariate or multivariate, they can be
summarize potential knowledge from the rich data broken up into univariate time series and univariate
[1]. A major part of these details is the time series. time series.
The so-called time series is a sequence of
chronologically sequential occurrences. With the use Anomaly detection challenge for time series is simply
of time series, scientists can accumulate and interpret achieved by finding outlier signal relative to some
a vast volume of data and understand more about regular signal, like unpredicted spikes, drops,
changes in trends and changes in level.

1895
Syed Hassan Ali Shah et al., International Journal of Advanced Trends in Computer Science and Engineering, 10(3), May - June 2021, 1895 – 1900

3. METHODS FOR ANOMALY DETECTION convolutionary decoder which detect and diagnostic
the remaining irregularities by using inter-sensor
Several foreign researchers have penetrated into the similitudes and the tempo information maps.
analysis of time series outlier detection after Barnett Extensive observation - based on virtual dataset and
wrote the first book Outliers of observational data on an existing plant data set experiment shows that
anomaly detection in the 1980s, such as M Breunig, MSCRED can use additional simple methods.
E M Knorr, E Keogh, Portnoy, J Takeuchi, M
Agyemang, M Markou, V Chandola and so on. For the issue of detecting and diagnosing
Domestic research begins very late, but advances abnormalities, they proposed it in that article, and
rapidly. Related research are being carried out by created a groundbreaking approach, MSCRED,
Tsinghua University, Xi'an Jiao Tong University, which integrates model reasoning. Multi-scale
Tianjin University, Fudan University, Hong Kong (resolution) device signature matrices are used to
University of Science and Technology, etc. Because characterize the state of the entire system at various
of the scientific relevance and deployment prospect time segments, and a deep encoder-decoder
of time series outlier identification, a large number of framework is used to produce reconstructed signature
scientists have joined in its study. Many high-quality matrices. The system is able to model both inter-
UTS achievements have been reached and published sensor associations and temporal dependencies in
in journals such as IEEE TKDE Neural Computing multivariate time series. After the residual signature
Numerical Statistics and Data Processing at a matrices have been extracted, they are further used to
renowned international conference such as PAKDD identify and diagnose any abnormalities. In a
PKDD SIGKDD VLDB over the past 10 years. comprehensive series of observational tests, which
compared the output of MSCRED on synthetic data
Anomaly detection is attracting even more and a power plant dataset, it was noticed that
recognition and analysis as an essential sub-branch of MSCRED outperforms the industry norms by a large
data mining. Most approaches of anomaly detection margin [3].
have been suggested by domestic and international
researchers, which can be classified into five 3.2. Unsupervised Anomaly Detection Using
categories: abnormal statistical-based detection, LSTM-Based Auto encoders
abnormal clustering-based detection, abnormal
distance-based detection, abnormal density-based A method for categorizing and detailing anomalies in
detection, etc. [2]. data sets has been identified as an anomaly. Correct
identification of anomalies today is crucial because
3.1. Deep Neural Network for Unsupervised pure data volumes prevent the hand-marking of
Anomaly Detection outliers. Auto detection system operations include
theft detection, physician supervision, error detection
Multi-turn data are currently gradually gathered from and incident detection. In this topic the main issue is
different real-time applications, such as power plants, that there are no anomalies. Therefore, conventional
wearable devices, etc. The multivariate pathological methods of machine learning cannot be used for
identifying and assessing sequence identifies and model training because time series labels are
separates at some stages the root causes of sporadic impossible.
illness. But it is important to design this mechanism,
not just to document time dependence and time Many other classical anomaly detection systems
series, but also to encrypt interrelationships between exist, for example
different time series pairs. The device may also be
noise tolerant and give operators numerous anomalies  Anomaly Detection Based on Clustering
depending on many collisions. Although many  Isolation Forests
unexpected anomaly detection algorithms have been  Support Vector Machine
developed, few can jointly solve these difficulties.  The application of Gaussian distribution for
We give a multi-variable time series for the anomaly detection
identification of deviations, a CRD in the article
(MSCRED). The multi-scaled (resolution) signature Moreover, all these methods describe the outlier
matrices MscrED originally produces to characterize merely because of its magnitude, but not because of
system status levels in various timescales. The the values of previous stages. In most other respects,
signature measurements then encode the associations when using such methods, temporal data is not taken
between the (time series) sensor and the attention- into account. Consequently, classical systems had
based CTM networks for transient model capture by little success. Irregularities in time series data can be
using a convolutionary encoder (ConvLSTM). mentioned among the observed algorithms.
Finally, an input signature matrix recreates the

1896
Syed Hassan Ali Shah et al., International Journal of Advanced Trends in Computer Science and Engineering, 10(3), May - June 2021, 1895 – 1900

 Outlier detection based on signal  Window scale for neural network feeding
decomposition (classical decomposition,  Window scale smoothing
STL) [4]  Threshold, where we regard as anomaly the
 Space vector model, Holt Winters, ARIMA height of residues
Exponentially smoothing
 Deep learning: auto encoders based on feed Other approaches may be helpful, but only the auto
forward, recurrent and LSTM neural encoder technique is universal and efficient enough
network layers [5] for all sorts of time series [6].
 Dimensionality reduction: RPCA, SOM,
discords, piecewise linear 3.3. Unsupervised Anomaly Detection with
LSTM Neural Networks
The easiest way to achieve this is to break down
during normal time series: seasonal components and Examine the detection of anomalies inside an
designs can first be excluded from the signal, along unattended framework and integrate long-term
with an outside traditional tracking system. For network neural memory (LSTM). These sequences
example, this approach works well on hotel price data are transferred in particular via our LSTM frame and
where constant fluctuations occur per year. Prices are achieve defined period in the specified sequences of
now that year-on-year due to inflation. variable data lengths. Then, you will notice our
anomaly detector's decision-making feature based on
So, we should look at the data point after being OC SVM and the single-class help-vector-definition
removed from the seasonal and the pattern vectors if (SVDD) algorithms. Because our first solution to
it is far from empty. However, where time series are collaborative preparation and optimization of LSTM
not accurate (e.g., foreign trade or sound), machine and OC-SVM algorithms is to use extremely
learning methods can only be employed. The self- effective gradients and quadratic programming. In
coding discovery of anomalies is one of the better order to incorporate gradient training methods, we
machine learning approaches. An automobile encoder modify the original objective criteria of OC-SVM
is an artificial nerve network used to encrypt data and SVDD algorithms if the current objective criteria
effectively and without monitoring. An encoder's converge with the original criteria. Our unattended
objective is to learn representation by training a formulation is often applied to semi-controlled and
network to disregard the signal 'noise' to reduce professionally supervised processes. It helps us
dimensionality for a wide range of data. achieve algorithms for anomaly detection that can
process and deliver high efficiency data sequences,
A rebuild and a reduction side are developed where particularly in time series data. Our approach is
the auto encoder plans to produce the same image as generic enough that our LSTM and GRU-based
the original input and term of the reduced encoding. architecture can be directly substituted with the Gated
A single encoder and a decoder layer also have Recurrent Units approach (GRU). In our research,
automated encoders, while profound encoders and our conventional algorithms show substantial
decoders are profit-making. The encoder and decoder increases in performance.
are used as two units. The features behind a stage are
identified by an encoder. These features are typically Anomaly detection is researched and LSTM
smaller. The decoder reconstructs the original data algorithms are presented in a non-supervised setting.
from these. Especially for the processing of variable-long data
sequences, we implemented a general LSTM
Feed forward’s neural network can be used to build framework. Following the acquirement of defined
auto encoder. We will therefore construct an LSTM- sequences via our LSTM-based architecture, we add
centric auto-encoder to accept temporal details. In a ranking feature of our OC-SVM [6] and SVDD [7]
comparison to a neural feed forward network, we use algorithms for anomaly detectors. The parameters of
information to refer to LSTM one at a time. Each both LSTM architectures and the final scoring
RNN unit is an extension of the RNN to preserve function for the OC-SVM (or SVDD) formulation are
awareness of its importance in the neural network in optimized as a first time in literature. We have also
a time sequence. conducted regression and Quadratic Programming-
based training sessions with various algorithmic
It is best to pick a neural network's design and post- values to refine the parameters for our algorithms
processing variables based on data to be fed into the together, so that our derivatives for these algorithms
device. The most relevant things to remember are: can be applied to the half-checked and totally
 Neural network's range of layers regulated frameworks. We change the OC-SVM and
 Layer LSTM cell count SVDD formulations in order to implement the
gradient-based training mechanism and then include

1897
Syed Hassan Ali Shah et al., International Journal of Advanced Trends in Computer Science and Engineering, 10(3), May - June 2021, 1895 – 1900

the convergence effects of the revised formulations properties in multiple applications. We build a
with the original formulations. Therefore, we get detection service in Microsoft which tracks millions
highly efficient anomaly detection algorithms, of measures from Bing, Workplace and Azure, and
specifically for time series data that can process data helps engineers function faster on a web. In this post,
sequences of varying lengths. We also have GRU- we emphasize the pipelines and the algorithm of our
based anomaly detection algorithms in our time series anomaly detection method.
simulations owing to the generic structure of our
method. We demonstrate major performance The system contains three core elements: data ingest,
improvements obtained with the traditional methods platform testing and online estimates. We will install
through our algorithms [7], [8] and [9] through a the whole pipeline first before describing these
broad variety of actual and virtual data sets via elements. By ingesting time series on a device, users
comprehensive experiments. will report monitoring incidents. It facilitates the use
of time series from different data sources (including
azure storage, databases, and online streaming data).
3.4. Time-Series Anomaly Detection Service at The ingestion manager shall vary with the
Microsoft granularities indicated at each stage, e.g., minute,
hour or day. Series points are stored in a time series
Large businesses must monitor their software and database on the streaming pipeline via Kafka. Online
facilities via various indicators in real time (e.g., state input time series anomaly test anomaly detector
website views and revenues). With Microsoft time processor. Consumers concurrently eat a number of
series, we provide an anomaly detector service that time series in a typical situation of demand
allows customers to monitor time series on a measurements. For example, for different markets
permanent basis and alert against events in time. We and channels, the Bing squad used a time series. If an
present in this text the pipeline and algorithm of our event occurs, alert systems combine time series
Anormal Detection Service for the unique, efficient anomalies, and provide email and payment services
and general purposes. The pipeline comprises three to customers. Cumulative abnormalities mean the
main components, namely intake of data, analysis average condition of an injury and allow users to
tools and online computing. This method aims to reduce diagnostic problems.
resolve the problem of identification of an anomaly
in time series by constructing a new algorithm based Anomaly identification in time series is important for
on Spectral Resin (SR) and CNN (CNN). Our work maintaining the consistency of online services. In real
was the first effort to identify abnormalities for time applications an inexpensive, robust and reliable
series by taking the SR model from the field of visual anomaly detection process is useful. We also released
saliency detection. We also merge SR with CNN for an Anomaly Detecting Service at Microsoft in this
innovative enhancement of the performance of SR article. More than 200 teams, among them Bing,
models. Our approach provides superior experimental Office and Azure, have used the service in Microsoft.
results, unlike the existing baselines on both the Anomalies in the production are detected from a
public and Microsoft data output platforms. maximum of 4 million time series per minute.
Moreover, for the first time in time series anomaly
The aim of identification of anomalies is to identify detection we apply the Spectral Residual (SR) model
unusual patterns or uncommon artifacts. Data mining and innovatively merge the SR model with the CNN
has become and is a critical analytical area for the model to deliver excellent performance. In future, we
business application as one of the most popular expect to combine together in order to provide our
sectors. The real time identity anomaly will save a customers with a better anomaly detection service.
company's resources by eliminating downtime, Besides internal service, as part of our Cognitive
mitigating brand damage and maintaining the Service, our time series anomaly detection system
company's image unchanged. Criminal justice experts will soon be accessible to external customers through
conclude that inaccurate, positive claims are Microsoft Azure [11].
primarily liable for regulation that, along with
financial companies who offer their own AMSI
services to manage the condition, commodities and 3.5. Deep Learning Approach for
the wellbeing of their industry, pose the greater Unsupervised Outlier Detection in Time
burden of this problem. When anomalies are Series
detected, administrators are advised that they should
take action to cope with injuries as soon as possible. In standard anomaly detection techniques, current
Yahoo's release of EGADS [10] is an excellent points and seasonal fluctuations normally found in
example, which aims to track and boost alerts for the streaming data cannot be observed on the basis of
millions of time series for Yahoo's numerous distances and density, thereby allowing temporal

1898
Syed Hassan Ali Shah et al., International Journal of Advanced Trends in Computer Science and Engineering, 10(3), May - June 2021, 1895 – 1900

anomalies to be identified in the existing IoT cycle. [15], but we use CNN for time series regression (and
We use a new approach to time series data to solve LSTM for comparison).
this problem, called the Deep Learning Detection
approach, which is applicable for non-streaming In cases where a significant number of data is
scenarios (DeepAnT). DeepAnT is able to detect a accessible without the risk of naming this method
vast variety of anomalies, including dots, background will practically be implemented. The data modeling
and time series discord. In comparison to methods of process may therefore be hampered by low data
anomaly identification under which anomalies are quality. In the other hand, if the amount of pollution
detected, DeepAnT uses uncontrolled data to gather is above 5%, the device will attempt to model the
and understand the dissemination of information used instances, then, as they are assumed to be natural at
for deterring natural behavior. DeepAnT has two the time of delimitation. The network design is
components: the time series predictor and chosen and the resulting hyper parameters are another
abnormality detector. The time series Modulator constraint. The modern architecture quest techniques
predicts when the defined horizon is next marked by [16] are the usage of human technological expertise
the profound neural networks (CNN). This module for this function to be circumvented. One of the most
takes a window with a time series and measures the serious constraints is perhaps the adverse examples
next time line (used as a background). The forecast [17] which restrict the use of this technique in
value is then transferred to the detector module, protection scenarios (and the majority of previous
which marks the time stamp continuously or data-driven methods). In knowing and defending
irregularly. Even without removing exceptions from against these adversary examples important measures
the data collection, DeepAnT can be trained. In have been made. But no generic strategy to overcome
general, for the deep learning methods of a model, this problem has yet been created [18].
several data are required. DeepAnT can acquire very
small data sets, but the close exchange of CNN
parameters guarantees a strong generalization power. 4. CONCLUSION
As DeepAnT does not recognize the anomalies, it
does not depend on unusual model generation marks. At the conclusion of this paper, a few different
This technique can also be used specifically in real- hypotheses on time series anomaly detection are
life situations, where a large amount of data from offered. The anomaly is just making a difference
heterogeneous sensors in natural and anomalous proportional to context. If a distinction is drawn
areas may be hard to distinguish. In addition to 10 between natural and abnormal behavior, it is
anomaly detection bench marks, we have conducted a pointless to regard one as another. The definition and
detailed evaluation of 15 algorithms, consisting of meaning of anomaly vary, depending on where one is
433 actual and synthetic time series. Experiments in the application. Because of this, say, the height of
show that DeepAnT is comparable with other a 568 cm isn't deemed out of the ordinary when it is
anomaly detection techniques in most situations. applied to a person, but for a CEO earning vastly
more than the rest of the workers, it is when a
The proposed DeepAnT comprises two modules. The comparison is drawn over a group. Therefore,
first module is the Time Series Index. The second specialists on the Internet programming deal with the
module is liable for the normal or abnormal marking reasons would influence the truthfulness of the
of data points in a given amount of occasions. Time- algorithm's results in determining whether or not the
scale is the second module. Deep learning was used algorithm detects accurately. To inform users of just
primarily for a vast range of applications because of the questionable or odd data in order to grab their
its potential to automatically discover complex attention Here we have an example of an extra or
features without domain knowledge. This artificial different condition, then viewed in relation to the
neural network learning ability enables a powerful data, we may call it an anomaly detection. Since
anomaly detection nominee in time series. Therefore standard multivariate data processing and
DeepAnT utilizes raw figures, which includes CNN. multidimensional expansion, the results of the MTS
It is often strong to change, in contrast to other neural seem to be very different. The factors have several
networks and mathematical models. Literature interrelationships; because of this, the results would
[12][13] is considered to function well with the likely have a multivariable structure. As there are
ability of LSTM to derive long-term trends from many possible reasons for detecting the MTS
time-series. However, we have seen that CNN can be variation, the MTS anomaly test is conducted by
an excellent alternative to standardized and multi- doing a systematic study of each component.
varying time series results because of its parameter Specifically, time series detection is still in MTS
performance. CNN and LSTM are generally used in anomaly remains immature, especially in MTS
literature classification issues of time series [14] and anomaly detection of deviance. In addition, the
existing anomaly detection algorithm is not

1899
Syed Hassan Ali Shah et al., International Journal of Advanced Trends in Computer Science and Engineering, 10(3), May - June 2021, 1895 – 1900

streamlined, so more reduction in complexity is 45–66, Jan. 2004. doi: 10.1023/

needed in order to cope with dynamic time series B:MACH.0000008084.60811.49.
analysis is still needs to be investigated. [9] P. Malhotra, L. Vig, G. Shroff, and P. Agarwal,
Long Short Term Memory Networks for Anomaly
ACKNOWLEDGE Detection in Time Series. Louvain-la-Neuve,
Belgium: Presses Universitaires de Louvain, 2015, p.
We thank Mr. Usman Ahmed Raza, our faculty 89
member, who encouraged us and supported us to [10] Nikolay Laptev, Saeed Amizadeh, and Ian Flint.
achieve this goal. 2015. Generic and Scalable Framework for
Automated Time-series Anomaly Detection. In
Proceedings of the 21th ACM SIGKDD International
Conference on Knowledge Discovery and Data
REFERENCES
Mining. ACM, New York, NY, USA, 1939–1947
[1] Bowen. Z. and Shahriar. S, "Finding needle in a [11] Hansheng Ren, Bixiong Xu, Yujing Wang, Chao
million metrics: anomaly detection in a large-scale Yi, Congrui Huang, Xiaoyu Kou∗ Tony Xing, Mao
computational advertising platform", Proceeding of Yang, Jie Tong, Qi Zhang. Time-Series Anomaly
the 2nd International Workshop on Ad Targeting at Detection Service at Microsoft. arXiv:1906.03821v1
Scale, San Francisco, CA, USA, pp.1062-1065, Feb, [cs.LG] 10 Jun 2019
2016 [12] F. A. Gers, D. Eck, and J. Schmidhuber,
[2] Ozkan, Huseyin.O, Fatih. O, Suleyman. S, ‘‘Applying LSTM to time series predictable through
"Online anomaly detection under markov statistics time-window approaches,’’ in Neural Nets WIRN
with controllable Type-I error ", IEEE Transactions Vietri-01. London, U.K.: Springer, 2002, pp. 193–
on Signal Processing, Vol.64, No.6, pp.1435-1445, 200
March, 2016. [13] S. Hochreiter and J. Schmidhuber, ‘‘Long short-
[3] Chuxu Zhang, ∗Dongjin Song, Yuncong Chen, term memory,’’ Neural Comput., vol. 9, no. 8, pp.
Xinyang Feng, Cristian Lumezanu, Wei Cheng, 1735–1780, 1997
Jingchao Ni, Bo Zong, Haifeng Chen, Nitesh V. [14] Z. C. Lipton, D. C. Kale, C. Elkan, and R.
Chawla. A Deep Neural Network for Unsupervised Wetzell. (2015). ‘‘Learning to diagnose with LSTM
Anomaly Detection and Diagnosis in Multivariate recurrent neural networks.’’ [Online]. Available:
Time Series Data, The Thirty-Third AAAI https://arxiv.org/abs/1511.03677
Conference on Artificial Intelligence, 2019. [15] Y. Zheng, Q. Liu, E. Chen, Y. Ge, and J. L.
[4] ] Robert B. Cleveland, William S. Cleveland, Zhao, ‘‘Time series classification using multi-
Jean E. McRae, Irma Terpenning. STL: A seasonal- channels deep convolutional neural networks,’’ in
trend decomposition procedure based on loss. Journal Proc. Int. Conf. Web-Age Inf. Manage. Cham,
of Official Statistics, Vol 6., No. 1, 1990, pp. 3-73. Switzerland: Springer, 2014, pp. 298–310.
http://www.nniiem.ru/file/news/2016/stl-statistical- [16] B. Zoph and Q. V. Le, ‘‘Neural architecture
model.pdf search with reinforcement learning,’’ CoRR, vol.
[5] A Tutorial on Deep Learning Part 2: abs/1611.01578, pp. 1–16, Nov. 2016.
Autoencoders, Convolutional Neural Networks and [17] A. Kurakin, I. J. Goodfellow, and S. Bengio,
Recurrent Neural Networks ‘‘Adversarial machine learning at scale,’’ CoRR, vol.
http://robotics.stanford.edu/~quocle/tutorial2.pdf abs/1611.01236, pp. 1–17, Nov. 2016.
[6] Oleksandr I. Provotar, Yaroslav M. Linder, [18] Mohsin Munir, Shoaib Ahmed Siddiqui,
Maksym M. Veres. Unsupervised Anomaly Detection Andreas Dengel, Sheraz Ahmed. DeepAnT: A Deep
in Time Series Using LSTM-Based Autoencoders Learning Approach for Unsupervised Anomaly
IEEE 2019 Detection in Time Series. IEEE January 7, 2019.
[7] ] B. Schölkopf, J. C. Platt, J. Shawe-Taylor, A. J. [19]AzlizaYacob ,ZirawaniBaharum , NurSukinah
Smola, and R. C. Williamson, “Estimating the Aziz, Noor SuhanaSulaiman and Wan Mohd Amir
support of a high-dimensional distribution,” Neural Fazamin Wan Hamzah, “A Review of Internet of
Comput., vol. 13, no. 7, pp. 1443–1471, 2001. Things (IoT): Implementations and Challenges”,
[8] ] D. M. J. Tax and R. P. W. Duin, “Support vector International Journal of Advanced Trends in
data description,” Mach. Learn., vol. 54, no. 1, pp. Computer Science and Engineering, 9(1.3), 2020,
373 - 376

1900