International Journal of Advanced Engineering, Management and Science (IJAEMS) [Vol-6, Issue-12, Dec-2020]

https://dx.doi.org/10.22161/ijaems.612.13 ISSN: 2454-1311

An Analysis of Outlier Detection through

clustering method
T. Chandrakala1, S. Nirmala Sugirtha Rajini2

1Assistant Professor, Department of Computer Applications, Jawahar Science College, Neyveli, TamilNadu, India
2Professor, Department of Computer Applications, Dr. M.G.R. Educational & Research Institute, Chennai, TamilNadu, India

Received: 08 Nov 2020; Received in revised form: 03 Dec 2020; Accepted: 12 Dec 2020; Available online: 30 Dec 2020
©2020 The Author(s). Published by Infogain Publication. This is an open access article under the CC BY license
(https://creativecommons.org/licenses/by/4.0/).

Abstract— This research paper deals with an outlier which is known as an unusual behavior of any substance
present in the spot. This is a detection process that can be employed for both anomaly detection and abnormal
observation. This can be obtained through other members who belong to that data set. The deviation present in
the outlier process can be attained by measuring certain terms like range, size, activity, etc. By detecting
outlier one can easily reject the negativity present in the field. For instance, in healthcare, the health condition
of a person can be determined through his latest health report or his regular activity. When found the person
being inactive there may be a chance for that person to be sick. Many approaches have been used in this
research paper for detecting outliers. The approaches used in this research are 1) Centroid based approach
based on K-Means and Hierarchical Clustering algorithm and 2) through Clustering based approach. This
approach may help in detecting outlier by grouping all similar elements in the same group. For grouping, the
elements clustering method paves a way for it. This research paper will be based on the above mentioned 2
approaches.
Keywords— detection of an outlier, data set, clustering approach, abnormality.

I. INTRODUCTION any data analysis is performed on the data(Bhattacharya et

Mining, in general, is termed as the intrinsic methodology of al., 2015).
discovering interesting, formerly unknown data patterns. A large number of domains apply Outlier detection directly.
Outlier detection has important applications in the field of This results in the development of innumerable outlier
data mining, such as fraud detection, customer behavior detection techniques. A lot of these techniques have been
analysis, and intrusion detection. A number of approaches developed to solve focused problems pertaining to a
are used in the process of detecting the outlier (Bezerra et particular application domain, while others have been
al., 2016). Clustering can be termed as a set-grouping task developed in a more generic fashion(Pimentel et al., 2014).
where similar objects are being grouped together. Clustering,
a primitive anthropological method is a vital method in
II. DATA IN DATA MINING
exploratory data mining for statistical data analysis, machine
Generally, we are drowned in information, but starving
learning, and image analysis and in many other predominant
for knowledge. Data can be collected from multiple sources.
branches of supervised and unsupervised learning.
The purposes can be categorized as Business, Science, and
Outlier detection is related to unwanted noise in the data. As Society. Business purposes can use the data for Web, E-
far as the analysts are concerned, noise in data is not Com, Transaction, and Stock Marketing. Scientific data can
important but acts as a hindrance to data analysis. Noise be used for Remote sensing, Bio-informatics, and Scientific
removal is the process of removing unwanted objects before

www.ijaems.com Page | 571

International Journal of Advanced Engineering, Management and Science (IJAEMS) [Vol-6, Issue-12, Dec-2020]
https://dx.doi.org/10.22161/ijaems.612.13 ISSN: 2454-1311

Simulation. Social data can be used for News, DigiCam, and Learning Patterns
YouTube. In data mining, how we now had to split the instruction
process into two classes particularly, Supervised
understanding and Unsupervised understanding. Back in
Supervised mastering, it's a kind of procedure by which the
two entered and desired output numbers have been all
provided. Input and output signal numbers are tagged for
classification to supply an understanding foundation for
prospective data processing system systems. The word
tricked learning stems out of the notion an algorithm is
learning more by your training data set, which is looked at
whilst an educator. The calculations include underneath
Supervised finding out are Conclusion bushes, Similarity
finding out, Bayesian logic, Service vector machines (SVM).
Back in Unsupervised mastering, there's not any requirement
• Data: a collection of facts usually obtained as the result to oversee this version. As an alternative, the version is
of experiences, observations, or experiments permitted to get the job done in its to detect advice. It largely
• Data may consist of numbers, words, images, … handles all the unlabelled info. Unsupervised learning
• Data: lowest level of abstraction calculations enable us to do more intricate processing
responsibilities in comparison to learning. Even though,
Data mining Techniques:
experiential learning could be unpredictable compared with
additional all-natural learning procedures. In this paper, we
focus on Hierarchical Clustering and also Kmeans
Clustering(Arthur & Vassilvitskii, 2006).
Clustering:
Having similar faculties clusters objects need to shape, using
the automatic procedure. We utilize clustering, to specify
lessons. Then suitable things have to place in each class.
Cluster calculations could be categorized based on the
cluster models readily offered based on the type of info
people we try to analyze [2]. In machine learning,
perspective clusters correspond to both hidden patterns, the
search for clusters is still unsupervised understanding, and
the resulting system represents an information concept. From
a practical standpoint, clustering Has an extraordinary role in
data mining programs like scientific information exploration,
There are six such data mining techniques namely Decision information retrieval, and text mining, spatial database
Trees, Sequential Pattern s, Clustering, Prediction, applications, Net analysis, CRM, promotion, health
Association, and Classification, out of which we deal with diagnostics, computational biology, and Others(Manning et
Clustering. al., 2008).

www.ijaems.com Page | 572

International Journal of Advanced Engineering, Management and Science (IJAEMS) [Vol-6, Issue-12, Dec-2020]
https://dx.doi.org/10.22161/ijaems.612.13 ISSN: 2454-1311

tagsas "ordinary" or even"outlier" usually are perhaps not

introduced. Clustering signifies learning observation as
opposed to learning samples. Clustering established an
outlier detection method for expanding information flow
that allots burden to feature based on its own significance
in mining endeavor(Manning et al., 2008).
The outlier detection procedure is also rather effective while
the info out of your database has been segmented into
clusters. In most bunch just about every data,the stage is
called a qualification of their registration. Even the outlier is
discovered with no hindrance from the clustering system.
Clustering on flowing information is distinguished by grid
established and kmeans/k median System(Xu & Wunsch,
2005)].
Hierarchical Clustering
A hierarchical clustering system operates by grouping info
items to some tree of clusters. The standard of the pristine
hierarchical clustering system is affected by the own
inability to do alteration the moment a mix or divide decision
was implemented. In other words, when your special unify
or divide decision afterward ends up to have now already
been a lousy option, then the procedure can't backtrack and
fix it. In hierarchical clustering delegate every single and
every item (info stage) into your bunch. Subsequently,
Cluster formation mechanism(Xu & Wunsch, 2005) calculate the length (correlation) among every one of the
clusters and then combine both similar clusters. Let us know
further by resolving a good model.
Cluster-based Ways for outlier detection in Statistics Sets
assist them to develop a set of equal aspects or bunch of
information details. Clustering methods are tremendously Dendrogram
Helpful for grouping related information objects from Data Objective: For the one-dimensional data set
sets and following this by employing space predicated {15,20,30,50,63}, perform hierarchical clustering and plot the
calculations, detection of Outliers has been accomplished, dendogram to visualize it.
which truly have been termed cluster-based outlier
Solution: First, let’s visualize the data.
detection.
The reward of this bunch of established approaches is they
usually would not need to become tricked. Moreover,
clustering established approaches have been capable to be
found in an incremental manner i.e later after learning how
the clusters, so fresh things might be fed into the machine
and also analyzed for outliers. 1 drawback of clustering
established procedures is They Are computationally costly
since they demand computation of pairwise
distances(Bhattacharya et al., 2015).
Clustering established outlier detection is also an
unsupervised outlier detection procedure at which category

www.ijaems.com Page | 573

International Journal of Advanced Engineering, Management and Science (IJAEMS) [Vol-6, Issue-12, Dec-2020]
https://dx.doi.org/10.22161/ijaems.612.13 ISSN: 2454-1311

Observing the plot above, we can intuitively conclude that: Hierarchical clustering is mostly used when the application
1. The first two points (15 and 20) are close to each other requires a hierarchy, e.g creation of a taxonomy. However,
and should be in the same cluster they are expensive in terms of their computational and
storage requirements.
2. The third point (30) is closer to the first formed
cluster(15, 20). So it is merged with the first cluster Agglomerative Hierarchical Clustering
next. This bottom-up strategy starts by placing each object in
3. Now the newly formed cluster is (15, 20, 30). its own cluster and then merges these atomic clusters into
larger and larger clusters until all of the objects are in a
4. Then, the next closer clusters are 50 and 63. So merge
single cluster or until certain termination conditions are
them in the next step, i.e. (50, 63).
satisfied. Most hierarchical clustering methods belong to this
Merge, in each step, the two clusters, whose two closest category. They differ only in their definition of inter-cluster
members have the smallest distance. similarity.

The related algorithm is shown below :

Given :
A set of X objects{X1,…Xn}
A distance function dist(c1,c2)
for I = 1 to n
Ci = {Xi}
end for
C = {c1,…,cn}
I = n+1
whilec.size> 1 do
(Cmin1,cmin2)=minimum dist(ci,cj) for all ci, cj in C
Remove cmin1and cmin2 form C
Two clusters formed are :
Add {cmin1, cmin2} to C
Cluster 1 : (15,20,30)
I = i+1
Cluster 2 : (50,63)
end while

www.ijaems.com Page | 574

International Journal of Advanced Engineering, Management and Science (IJAEMS) [Vol-6, Issue-12, Dec-2020]
https://dx.doi.org/10.22161/ijaems.612.13 ISSN: 2454-1311

K-Means Clustering centroids. 1 closing Thing to stay in your mind is how k

Hierarchical clustering is most beneficial at detecting means is intended to use on steady statistics. You will find
embedded buildings inside the info. But it neglects in locating plenty of implementations to get K Means clustering
aconsensus' round the complete data set. Hierarchical accessible on the market, A Few Of these are Apache
clustering can place with clusters that seem shut, however, no Mahout,
more advice concerning additional things can be thought. K A Far More Comprehensive collection of software that uses
means procedure simply consider a little neighborhood of outlier detection is:
points that are nearby and also additionally don't regard the 1. Fraud-detection - discovering deceptive applications
complete data set. for bank cards, say advantages or discovering fraud
At a feeling, k means believes just about each and every using charge cards or even cellular telephones.
single point from the data set and makes use of this advice to a. Advance application processing company - to -
evolve that the clustering above a succession of iterations. K find fraudulent software or maybe socialize
means has become easily the hottest clustering algorithm clients.
which reduces the total amount of their within-cluster b. Intrusion-detection - discovering rapid
variances. accessibility in pc system networks.
K means have plenty of variants to Boost for particular sorts c. Task tracking - discovering mobile-phone
of information. To a top degree, All of Them do something fraud by tracking cellphone exercise or
such as that: questionable transactions at the equity markets.
d. Network effectiveness - tracking the operation
K means selections points from multi-purpose distance to
of personal computer programs, such as to find
reflect every one of the clusters. All these are termed as
system bottlenecks.
centroids. Every individual will likely probably be nearest to
e. Fault investigation - tracking procedures to find
an inch of those bronchial centroids. They won't always be
flaws in engines, generators, pipelines, or
nearest to the exact same individual, therefore that they'll
distance tools on distance shuttles such as.
produce a bunch on their closest centroid. That which we are
f. Biomedical flaw detection - tracking producing
clusters and just about every affected person is presently an
lines to find faulty generation operations such
associate of the bunch. K means subsequently locates out the
as busted beams.
center for every one of the k clusters predicated on its own
g. Satellite picture evaluation - pinpointing
audience members (yep, employing the affected person
publication options or misclassified attributes.
vectors!). This center gets to be the brand's newest centroid
h. Discovering novelties in graphics - to get robot
for your own audience. Due to the fact the centroid is at
geotaxis or surveillance procedures.
another place today, sufferers could be more closer to
i. Movement segmentation - discovering image
additional centroids. To put it differently, they can change
includes relocating independently of this
audience membership. Duplicate before centroids no-longer
desktop.
shift, and also the bunch memberships stabilize. That really is
j. Timeseries tracking - tracking safety vital
known as convergence.
software like high-speed or drilling milling.
The essential selling purpose of k means is its own simplicity. k. Medical requirement tracking - including as for
Its ease means that it's generally speedier and better compared example for instance heartrate screens.
to many other calculations, notably during large data sets. It l. Pharmaceutical exploration - determining
becomes easier: k means may be utilized to pre-cluster that a publication molecular arrangements.
more gigantic data-set accompanied closely with way of an m. Discovering novelty in the text to find the
expensive audience examination around the sub-clusters. beginning of information reports, such as
Kmeans may likewise be utilized to immediately "drama" subject detection and monitoring to get dealers
with k and research if you will find missed styles or to directly successfully nail fairness, products,
connections inside the data set. foreign currency trading reports, out-
The flaws of K Means are its own significance to Outliers, performing or under-performing goods.
also its particular own sensitivity for the very first selection of

www.ijaems.com Page | 575

International Journal of Advanced Engineering, Management and Science (IJAEMS) [Vol-6, Issue-12, Dec-2020]
https://dx.doi.org/10.22161/ijaems.612.13 ISSN: 2454-1311

n. Discovering unexpected entrances in Data [10] Arthur, David, and Sergei Vassilvitskii. k-means++: The
Bases for information Mining to find glitches, advantages of careful seeding. Stanford, 2006.
valid or frauds however abrupt entrances(Xu & [11] Angelin, B., and A. Geetha. "Outlier Detection using
Clustering Techniques–K-means and K-median." 2020 4th
Wunsch, 2008).
International Conference on Intelligent Computing and
Control Systems (ICICCS). IEEE, 2020.
III. CONCLUSION [12] Angelin, B., and D. Devakumari MCA. "Outlier Detection
using Clustering Algorithm A Survey."
This research paper deals with the process of detecting
outlier through the clustering approach. Outlier which is
known as an unusual behavior of any substance present in
the spot. This is a detection process that can be employed for
both anomaly detection and abnormal observation. This can
be obtained through other members who belong to that data
set. The positives and negatives of dealing with K-Means
and Hierarchical Clustering have been discussed. The
algorithm must be modified in order to obtain proper results
for detecting outlier. Further study on this research will
concentrate more on enhancing the algorithm to obtain a
better result.

REFERENCES
[1] Benjelloun, Fatima-Zahra, Ayoub Ait Lahcen, and Samir
Belfkih. "Outlier detection techniques for big data streams:
focus on cybersecurity." International Journal of Internet
Technology and Secured Transactions 9.4 (2019): 446-474.
[2] Sualeh, Muhammad, and Gon-Woo Kim. "Dynamic multi-
lidar based multiple object detection and
tracking." Sensors 19.6 (2019): 1474.
[3] Zhao, Xingwang, et al. "Optimal design of an indoor
environment by the CFD-based adjoint method with area-
constrained topology and cluster analysis." Building and
Environment 138 (2018): 171-180.
[4] Wang, Zhao-Yu, et al. "Weighted z-Distance-Based
Clustering and Its Application to Time-Series Data." Applied
Sciences 9.24 (2019): 5469.
[5] Bezerra, Clauber Gomes, et al. "An evolving approach to
unsupervised and real-time fault detection in industrial
processes." Expert systems with applications 63 (2016): 134-
144.
[6] Bhattacharya, Gautam, Koushik Ghosh, and Ananda S.
Chowdhury. "Outlier detection using neighborhood rank
difference." Pattern Recognition Letters 60 (2015): 24-31.
[7] Pimentel, Marco AF, et al. "A review of novelty
detection." Signal Processing 99 (2014): 215-249.
[8] Manning, Christopher D., Hinrich Schütze, and Prabhakar
Raghavan. Introduction to information retrieval. Cambridge
university press, 2008.
[9] Xu, Rui, and Donald Wunsch. "Survey of clustering
algorithms." IEEE Transactions on neural networks 16.3
(2005): 645-678.

www.ijaems.com Page | 576