See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/368983958

An Unsupervised Machine Learning Algorithms: Comprehensive Review

Article in International Journal of Computing and Digital Systems · April 2023

DOI: 10.12785/ijcds/130172

CITATIONS READS
65 13,742

4 authors:

Samreen Naeem Aqib Ali

Islamia University of Bahawalpur Concordia College Bahawalpur
33 PUBLICATIONS 419 CITATIONS 51 PUBLICATIONS 688 CITATIONS

SEE PROFILE SEE PROFILE

Sania Anam Munawar Ahmed

Govt Associate College for Women Ahmadpur East, Bahawalpur, Pakistan Islamia University of Bahawalpur
32 PUBLICATIONS 473 CITATIONS 16 PUBLICATIONS 135 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Aqib Ali on 16 April 2023.

The user has requested enhancement of the downloaded file.

 International Journal of Computing and Digital Systems
ISSN (2210-142X)
Int. J. Com. Dig. Sys.13, No.1 (Apr-23)
http://dx.doi.org/10.12785/ijcds/130172

An Unsupervised Machine Learning Algorithms:

Comprehensive Review
Samreen Naeem1 , Aqib Ali1 , Sania Anam2 and Muhammad Munawar Ahmed3
1
College of Automation, Southeast University, Nanjing 210096, China.
2
Department of Computer Science, Govt Associate College for Women Ahmadpur East, Bahawalpur, Pakistan.
3
Department Information Technology, The Islamia University of Bahawalpur, Bahawalpur, Pakistan.

Received 25 May. 2022, Revised 19 Dec. 2022, Accepted 6 Feb. 2023, Published 16 Apr. 2023

Abstract: Machine learning (ML) is a data-driven strategy in which computers learn from data without human intervention. The
outstanding ML applications are used in a variety of areas. In ML, there are three types of learning problems: Supervised, Unsupervised,
and Semi-Supervised Learning. Examples of unsupervised learning techniques and algorithms include Apriori algorithm, ECLAT
algorithm, frequent pattern growth algorithm, clustering using k-means, principal components analysis. Objects are grouped based on
their same properties. The clustering algorithms are divided into two categories: hierarchical clustering and partition clustering. Many
unsupervised learning techniques and algorithms have been created during the last decade, and some of them are well-known and
commonly used unsupervised learning algorithms. Unsupervised learning approaches have seen a lot of success in disciplines including
machine vision, speech recognition, the creation of self-driving cars, and natural language processing. Unsupervised learning eliminates
the requirement for labeled data and human feature engineering, making standard machine learning approaches more flexible and
automated. Unsupervised learning is the topic of this survey report.

Keywords: Machine Learning, Unsupervised Learning, Clustering, Unsupervised Algorithms.

1. INTRODUCTION operate on particular data without external instruction. Un-

Because of its inherent concepts, machine learning can supervised learning approaches are essential to constructing
be considered a sub field of Artificial Intelligence (AI). human artificial intelligence systems [5]. Because intelligent
Allows for prediction; its core components are algorithms. machines must be able to make (independent) conclusions
ML enables systems to learn independently without having based on enormous amounts of unlabeled data, this is the
to be explicitly programmed to do so, resulting in more case. UL algorithms are better at solving complicated tasks
intelligent behavior. Develop models that detect trends in than supervised learning algorithms. However, supervised
historical data and utilize those models to make forecasts learning models provide more accurate results because a
to generate data-driven predictions [1], [2], [3]. Figure 1 de- programmer explicitly teaches the system what to search for
picts the overall architecture of machine learning, which is in the data presented. Unsupervised learning, on the other
made up of multiple steps: Modeling (which includes func- hand, may be highly unexpected. Unsupervised learning
tionality engineering, model training, and assessment), Data may be the foundation for artificial neural networks, making
Acquisition and Understanding (data collecting and under- deep learning possible [6]. While this is true, supervised
standing), and Business Comprehending (understanding and learning techniques for neural networks may also be used
understanding the domain) as well as a deployment (deploy if the intended output is already known. Learning without
the model to the cloud). Unsupervised Learning (UL) is a supervision may be a goal in and of itself. UL models, for
machine learning approach for detecting patterns in datasets example, may be used to uncover hidden models in massive
with unlabeled or unstructured data points. In this learning amounts of data and categorize and label data points. The
approach, an artificial intelligence system gets just the similarities and contrasts between unsorted data points are
input data and not the associated output data. Unsupervised used to group them [7].
machine learning, unlike supervised learning, does not need
the presence of a person to oversee the model [4]. The data Here are a few reasons why unsupervised learning
scientist enables the system to learn on its own by looking is so important.
at the data and identifying patterns. To put it another way,
this sub-category of machine learning allows a system to • There is a lot of unlabeled data.

http:// journals.uob.edu.bh
E-mail address: samreencsit@gmail.com, aqibcsit@gmail.com, sania.anam7@gmail.com, munawar69@iub.edu.pk
912 Samreen Naeem, et al.: An Unsupervised Machine Learning Algorithms: Comprehensive Review.

Figure 1. Basic Architecture of ML

• Data tagging is a time-consuming operation that ne- ture—this study surveys time-series clustering techniques.
cessitates human intervention. The uniqueness and limitations of earlier studies are also
explored, along with prospective research areas. Time series
• However, ML may be used to drive the same process, clustering applications are also listed. This literature review
making coding easier for everyone involved. focuses on time series clustering approaches. The reference
[11] surveys unsupervised and semi-supervised clustering
• It can be used to investigate unknown or unprocessed that describe clustering techniques and methodologies. The
data. authors gave external and internal clustering validity mea-
• It comes in handy when dealing with massive data sures. Their work helps researchers, although their literature
sets and pattern detection. review is limited to algorithms and clustering.
B. Motivation and Contribution
A. Literature Review
Unsupervised algorithms are extensively employed to
Many researchers have surveyed unsupervised learning
complete data mining jobs; they are discussed alone or
(UL) techniques. The reference [8] surveyed unsupervised
in groups based on learning needs. Literature studies on
learning literature. Their study included 49 studies. UL
supervised algorithms tend to focus little on unsupervised.
models are equivalent to supervised learning (SL) models,
The authors analyzed 35 papers between 2018 and 2022
and Fuzzy C-means and Fuzzy SOMs perform best among
and found that majority focused on unsupervised learning
UL methods. Their work focused on UL models for soft-
techniques. This review focused on unsupervised machine
ware fault prediction. The reference [4] analyzed supervised
learning techniques developed between 2018 and 2022.
and UL studies using a literature scan. They prioritized
research works published between 2015 and 2018 that 2. UNSUPERVISED LEARNING
address or use supervised and unsupervised ML approaches. In supervised learning, a data scientist offers labeled
This survey only included k-means, hierarchical cluster- data to the system, such as photographs of cats tagged
ing, and PCA. The reference [9] surveyed UL multiway as cats, so that it may learn by example. In unsupervised
models, algorithms, and their applications in chemometrics, learning, a data scientist merely gives photos, and it is up
neurology, social network analysis, text mining, and com- to the system to examine the data and determine whether
puter vision. Their poll exclusively analyzed unsupervised or not they are cat images. Large amounts of data are
multiway data. The reference [10] surveyed the litera- required for unsupervised machine learning [12]. In most
http:// journals.uob.edu.bh
 Int. J. Com. Dig. Sys. 13, No.1, 911-921 (Apr-23) 913

circumstances, supervised learning works similarly, with time [14], [15]. The workflow of the clustering algorithm
the model becoming more accurate as more examples are is shown in Figure 3.
added. When data scientists use datasets to train algorithms,
the unsupervised learning process begins. These datasets
include no labeled or classed data points. The purpose of
learning the algorithm is to find patterns in the dataset
and rate the data points according to those patterns. The
clustering, association, anomaly detection, and autoencoder
issues are four types of unsupervised learning challenges,
as shown in Figure 2.
Figure 3. Workflow of clustering in unsupervised ML

A. Types of Unsupervised Learning

1) Clustering
The practice of classifying items into groups is known
as clustering or cluster analysis. Clustering may be divided
into several forms, including partitioning, hierarchical, over-
lapping, and probabilistic. Data is partitioned so that each
piece of information may only belong to one cluster [16].
Exclusive pooling is another term for it. K-means exemplify
partitioning. Each data point is a cluster in hierarchical
clustering. The number of clusters is reduced via iterative
connections between the two closest clusters. It is used to
organize data in overlapping fuzzy sets [17], [18] . Figure
4 shows the example of clustering. Each point can be

Figure 2. Types of Unsupervised Learning

In the case of cat photos, the unsupervised learning

system may learn to recognize distinguishing features like
Figure 4. Example of Clustering
whiskers, long tails, and retractable claws. Unsupervised
learning is how humans learn to identify and classify objects
assigned to two or more categories with varying degrees of
they think about. Let’s pretend you’ve never had ketchup
affiliation. In this case, the data will be paired with a suitable
or spicy sauce before [13]. You will be able to discern the
membership value, such as K-Means clustering. Finally,
difference between two ”unlabeled” bottles of ketchup and
the probability distribution is used to generate clusters in
chili sauce if you are given two ”unlabeled” bottles of each
probabilistic [19]. Clustering may be divided into three
and asked to taste them. Even if you don’t know the names
categories according to how they operate, as shown in
of either sauce (one sour, the other spicy), you’ll be able
Figure 5.
to recognize its characteristics. You can get a better sense
of the flavor by tasting them more times. Soon, you’ll be Clustering may be divided into three categories accord-
able to categorize foods depending on the amount of sauce ing to how they operate, as shown in Figure 5.
they contain just by tasting them. Specific qualities that
distinguish the two sauces and the group’s nutrition can • Exclusive Clustering: As the name implies, complete
be discovered by studying the flavor. To categorize them, clustering states that a data item or object may only
you don’t need to know the names of the sauces or the exist within a cluster [14].
foods. Alternatively, you might refer to one as sweet sauce
and the other as a spicy sauce. Machines use unsupervised • Hierarchical Clustering: Hierarchical clustering [20]
learning to find patterns and similarly classify data points. seeks to establish a hierarchy of clusters. Agglomer-
In the same way, supervised learning would entail someone ated and divided hierarchical grouping are the two
informing you of the sauce names and flavors ahead of forms of the hierarchical set. Agglomeration uses a
http:// journals.uob.edu.bh
914 Samreen Naeem, et al.: An Unsupervised Machine Learning Algorithms: Comprehensive Review.

that has to be cleaned before analysis. When data models go

beyond or diverge from usual models, this is an anomaly. An
unusual network traffic pattern, for example, indicates that
the compromised system is transferring sensitive data to an
unauthorized server. Anomalies are identified or predicted
by finding or forecasting data points that differ from the
standard model [24], [25]. Intrusion detection, insurance,
fraud detection, and military surveillance are just a few of
the uses for anomaly detection.
4) Autoencoders
Autoencoders are an unsupervised learning approach
that uses neural networks to do representation learning.
We’ll create a neural network architecture with a bottleneck
that compels the network to use a compressed knowledge
representation of the original input [26]. This compression
and subsequent reconstruction would be complicated if the
input properties were unrelated. If the data has some struc-
ture (for example, correlations between input attributes),
this structure may be learned and utilized by driving the
Figure 5. Types of Clustering input through the network bottleneck. The presence of
an information bottleneck is a fundamental feature of our
network design; without it, our network might quickly learn
bottom-up method, first considering each data point to store input data by passing them through the web [27].
as a single group, then merging the pairs of groups Autoencoder algorithm components are shown in Figure 6.
as they advance up the hierarchy. Agglomerate is the
polar opposite of divisive. As you move down the
structure, each data point begins in a single group
and then divides.
Figure 6. Autoencoder Components
• Overlapping Clustering: You may use overlapping
grouping to divide a data point into two or more 3. REVIEW ON ML UNSUPERVISED ALGORITHM
groups [21]. We thoroughly analyze the literature on unsupervised
learning methodologies and algorithms and performance
2) Association measures used in unsupervised learning. The benefits and
The unsupervised learning approach of Association Rule drawbacks of various unsupervised learning research in this
Learning (ARL) is used to uncover associations between paper. This research will help academics go in a new direc-
variables in massive datasets. ARL can accept non-numeric tion by identifying new research areas and filling a research
data points, unlike specific machine learning methods. In vacuum in unsupervised learning [28]. Researchers will be
a nutshell, ARL is concerned with how particular variables able to compare the efficiency and efficacy of unsupervised
are linked. People who purchase a motorcycle, for example, learning algorithms as part of this project. Algorithms are
are more likely to get a helmet [22]. It’s possible to make used to implement both clustering and association rule
money by forming these kinds of connections. Suppose learning. Some of the most significant algorithms used to
consumers who purchase product X also purchase product apply the association rule include the a priori method, the
Y, and an online shop can suggest product Y to anybody ECLAT algorithm, and the frequent model growth (FP)
who buys product X. Internally. In that case, statements are algorithm. Algorithms like k-mean clustering and principal
used to learn the laws of association. These assertions may component analysis make clustering possible (PCA), as
highlight connections between disparate data sets. Support shown in Figure 7 [29], [30].
and trust are used if/when patterns or relationships are
identified. The media determines the frequency with which A. Apriori Algorithm
the if / then connection appears in the database. The number The Apriori algorithm was created with data mining in
of times the if/then the relationship has been determined mind. It may be used to extract data from databases with
to be legitimate is called confidence. The association rule many transactions, such as a database containing a list of
allows shopping cart analysis and online use mining [23]. things purchased by supermarket customers. It’s used to
detect the collection of things that customers are more likely
3) Anomaly Detection to buy together in shopping cart analysis and to identify the
Any procedure that discovers outliers in a data set adverse effects of medications [31].
is known as anomaly detection. These anomalies might
suggest unusual network activity, a faulty sensor, or data The stages for the Apriori algorithm are as follows:
http:// journals.uob.edu.bh
 Int. J. Com. Dig. Sys. 13, No.1, 911-921 (Apr-23) 915

• In the medical profession, for example, patient

database analysis.
• In forestry, data from forest fires is used to analyze
the frequency and intensity of forest fires.
• Many firms employ it a priori, such as Amazon in
their referral system and Google in their autocomplete
feature.

B. ECLAT Algorithm
ECLAT (Equivalence Class Clustering and Bottom-up
Lattice Traversal) is a data mining technique used to acquire
element set mining and locate frequent items. Because the
a priori technique utilizes a horizontal data structure, it
must scan the database numerous times to find frequently
occurring objects. On the other hand, ECLAT takes a
vertical approach and is faster in general since it only has
to scan the database once [35].

Figure 7. Widely used Unsupervised Learning Algorithms

• Step 1: Determine the transactional database’s sup-

port for the article sets and select the slightest degree
of support and reliability.
• Step 2: Select any available media with a more
excellent support value than the minimum or specified
support value.
• Step 3: Find all the rules in these subgroups with a
higher confidence value than the threshold or minimal
confidence.
• Step 4: Sort the rules in ascending order by their Figure 8. ECLAT Algorithm Model
elevation.
The stages for the ECLAT algorithm are as follows:
1) Advantages
• This is a straightforward algorithm. • Step 1: For each item in the database, get a list of
• On big datasets, the algorithm’s join and prune phases the transaction ID. We scan the entire database in this
are simple to implement [32]. step. The list of transactions containing element an is
the t transaction ID list of element a.
2) Disadvantages
• Step 2: Create a new list of transactions whose
• In comparison to other algorithms, the a priori algo-
members are transactions involving elements a and
rithm is sluggish.
b by intersecting the list of element titles a with the
• Because it checks the database many times, overall list of element titles b.
performance may suffer.
• Step 3: Apply conditional a to other items in the
• The a priori algorithm’s spatial and temporal com- database and repeat step 1.
plexity is O (2D), which is exceptionally high. The
• Step 4: Repeat the preceding procedures for the
horizontal breadth of the database is represented by
remaining items.
D [33].
3) Applications 1) Advantages
• Extraction of association rules in data mining of ad- • The Eclat method utilizes less memory than the Apri-
mitted students based on features and specializations ori algorithm since it uses a deep search technique
in the educational area [34]. [36].

http:// journals.uob.edu.bh
916 Samreen Naeem, et al.: An Unsupervised Machine Learning Algorithms: Comprehensive Review.

• Eclat’s approach does not require repetitive input • Step 6: The constructed FP tree must now be ex-
scanning to compute individual support values. tracted. The lowest node, as well as the relationships
between the weakest nodes, are evaluated first. The
• Unlike Apriori, which scans the original dataset, the lowest node represents the length of frequency model
Eclat algorithm searches the recently created datasets. 1. Then take the way via the FP tree. The conditional
model base refers to this path or pathways. The
2) Disadvantages dependent model is based on a secondary database
• The Eclat algorithm uses more RAM to build inter- containing prefix pathways in the FP tree that begin
mediate transaction ID sets. at the lowest node (suffix).
3) Applications • Step 7: Count the number of sets of items in the
• In the medical profession, for example, patient path to create a conditional FP tree. The hanging FP
database analysis. tree considers the collections of items that pass the
support criterion.
• In forestry, data from forest fires is used to analyze
the frequency and intensity of forest fires [36]. • Step 8: Create a conditional FP tree by counting the
number of sets of items in the route. The hanging
C. Frequent Pattern Growth Algorithm FP tree considers the groups of things that pass the
The Apriori algorithm has been improved with the support threshold.
Frequent Pattern (FP) Growth algorithm. This algorithm
• Step 9: The conditional FP tree generates frequent
represents the database in the form of a pattern or frequent
patterns.
tree (FT) structure. The most common patterns are extracted
using this regular tree. The Apriori technique must search 1) Advantages
the database n + 1 times (where n is the most extended
model’s length), but the FP growth approach needs two • This approach only needs to scan the database twice,
scans [37]. The stages for the Frequent pattern (FP) growth compared to Apriori, which examines the transactions
algorithm are as follows: for each iteration [38].
• This method avoids item matching, which speeds up
• Step 1: The first step is to run a database scan to the process.
see whether there are any occurrences of the item
sets. This is the same as the first step in the Apriori • Extraction of long and short frequent patterns is
method. The supporting count or frequency of one efficient and scalable since the database is compressed
collection of items in the database is the frequency in memory.
of one set of things.
2) Disadvantages
• Step 2: The FP tree is built. Begin by constructing
• FP Tree is bulkier and more complicated to build than
the tree’s root. The word null is used to represent the
Apriori, and it might be rather expensive.
root.
• The approach may not fit in shared memory if the
• Step 3: Re-scanning the database and going over all
database is extensive.
the transactions is the next stage. Examine the first
transaction to see what items it contains. The highest- 3) Applications
counting things are taken first, then the lowest-
counting items, etc. It denotes that the tree branch • Clustering, classification, software issue identifica-
is built up of sets of transaction components in tion, recommendations, and other problems may all
decreasing order of count. be solved with a Frequent pattern (FP) growth algo-
rithm [38].
• Step 4: The next transaction in the database is
examined. The object sets are ordered in ascending D. Clustering using K-Means
order by count. If a group of components from this In data science, several rounds of the k-means method
transaction already exists in the root, this transaction are commonly utilized. The k-means clustering algorithm
branch will share a common prefix. This signifies that divides components into groups based on their similarity.
the standard item set is linked to the new node of Graphically representation of the K-mean clustering work-
another item set in this transaction. flow is shown in Figure 9.

• Step 5: The item set count increases as transactions The letter k denotes the number of groups. As a result,
are made. As new nodes are established and joined if k is 3, there will be three groupings [39], [40], [41].
based on transactions, the count of both the familiar This clustering algorithm divides the unlabeled dataset into
and new nodes increases by one. unique clusters with comparable qualities for each data

http:// journals.uob.edu.bh
 Int. J. Com. Dig. Sys. 13, No.1, 911-921 (Apr-23) 917

• K-mean clustering provides smaller groups than hier-

archical clustering.
2) Disadvantages
• It’s impossible to determine the number of clusters or
the value of k.
• The number of groups in a network (k value) and
other initial inputs is strongly influenced by the
output.
• The order in which the data is input significantly
impacts the final result.
• It is pretty sensitive to changes in size. The outcome
will be dramatically different if we scale our data
Figure 9. Workflow of K-mean Clustering
using normalization or standards—the ultimate result
[42].
point. The trick is to locate cluster centroids, which are • If the nests have a complex geometric shape, nesting
K centers. Each group will have a cluster centroid. After is not suggested.
presenting a new data point, the algorithm will use metrics
like Euclidean distance to identify which cluster the data 3) Applications
point belongs to. The centroids are calculated using the • Segmentation of the market
K-mean clustering technique, which iterates until the best
centroid is obtained. The number of groupings is likely • Document grouping and picture segmentation
known. Another term for it is the flat clustering algorithm.
The number of groups found by the algorithm is denoted • Compression of images
by the letter ”K” in K-means [42].
• Segmentation of customers
The steps below will demonstrate how the K-Means
• Dynamic data trend analysis
clustering approach works:
E. Principal Components Analysis (PCA)
• Step 1: We must first specify the number of groups PCA is a dimensionality reduction approach that reduces
(K) that this method should produce. the dimensionality of big data sets by converting many vari-
ables into a smaller group that still maintains the majority of
• Step 2: Next, pick K random data points and assign
the data information in the ’large set.’ The loss of precision
them to one of the groups. In a nutshell, it ranks data
while lowering the number of variables in a dataset is
according to the number of data points it contains.
unavoidable, but the answer to reducing dimensionality is
• Step 3: Now, we’ll calculate the cluster centroids. to exchange some accuracy for simplicity. Because smaller
datasets are easier to study and visualize and there are no
• Step 4: Repeat steps 1-3 until you discover the unnecessary variables to analyze, data analysis is easier and
optimum centroid, i.e., allocating data points to non- faster for machine learning algorithms [43]. The graphical
varying groups. representation of the K-mean clustering workflow is shown
◦ 4.1: To begin, compute the total of the squared in Figure 9.
distances between the data points and the cen-
troids.
◦ 4.2: Now, we must allocate each data point to
the closest group to the others (centroid).
◦ 4.3: Finally, by averaging all data points in the
cluster, compute the centroids of the groups.
1) Advantages
• It is simple to comprehend and put into practice.
• If we had a high number of variables, K-mean would Figure 10. Workflow of PCA
be quicker than hierarchical clustering.
To summarize, PCA’s goal is to decrease the number
• When the centroids are recalculated, the cluster of an of variables in a data collection while retaining as much
instance might be modified. information as feasible.
http:// journals.uob.edu.bh
918 Samreen Naeem, et al.: An Unsupervised Machine Learning Algorithms: Comprehensive Review.

TABLE I. A Comparison of Five Commonly used Unsupervised Algorithms

Apriori Algorithm ECLAT Algorithm Frequent Pattern Growth Algorithm K-Means Algorithm Principal Components Analysis Algorithm
Accuracy in General Satisfactory Good Good Superb Superb
Speed of Learning Good Excellent Good Superb Superb
Speed of Classification Superb Superb Good Excellent Superb
Tolerance to Missing Values Good Excellent Superb Superb Superb
Tolerance to Irrelevant Attributes Good Excellent Satisfactory Superb Good
Tolerance to Redundant Attributes Excellent Good Satisfactory Excellent Good
Tolerance to Highly Interdependent Attributes Satisfactory Good Excellent Excellent Satisfactory
Tolerance to Noise Excellent Good Satisfactory Superb Superb
Dealing with Danger of Overfitting Good Good Superb Excellent Excellent
Attempts for Incremental Learning Superb Satisfactory Good Good Superb
Transparency of Knowledge/ Classification Satisfactory Good Excellent Excellent Good
Support Multi classification Good Excellent Excellent Superb Superb

The steps below will demonstrate how the PCA • The trade-off between dimensionality reduction and
approach works: information loss. Reduced dimensionality is benefi-
cial, but it comes at a cost. Information loss is an
• Step 1: Obtain the data set unavoidable aspect of the PCA [44].
• Step 2: Data representation in a structure. 3) Applications
• Step 3: Standardization of data • PCA is mainly utilized in artificial intelligence appli-
cations such as computer vision, image compression,
• Step 4: Z’s covariance is calculated. and a dimensionality reduction approach.

• Step 5: Eigenvalues and eigenvectors are calculated. • If the data is vast enough, it may also be utilized to
find hidden models. Finance, Data Mining, Psychol-
• Step 6: The eigenvectors are categorized. ogy, and more areas employ PCA [45].
• Step 7: New characteristics or primary components 4. COMPARATIVELY ANALYSIS
are calculated. So, here’s a comparison of the most popular unsuper-
• Step 8: Remove characteristics from the new dataset vised classification algorithms. Several strategies have been
that are less significant or irrelevant. created, some of which have been addressed in earlier
sections. Based on available facts and theoretical stud-
1) Advantages ies, Table 1 compares various regularly used unsupervised
• PCA helps us to better generalize machine learning algorithms. This comparison demonstrates that no single
models by lowering the dimensionality of the input. learning algorithm beats the others.
This aids us in overcoming the ”dimensionality curse”
[44]. 5. CONCLUSION
Unsupervised learning is one of the many types of ma-
• The calculation is simple. PCA is based on linear chine learning. The model is trained on an unlabeled dataset
algebra, which computers can solve quickly. in unsupervised learning. Grouping, association, anomaly
detection, and automated encoders are also included. Var-
• Other machine learning algorithms will be sped up. ious techniques for unsupervised learning have been pre-
Machine learning algorithms are trained on crucial sented throughout the last decade. Unsupervised learning
components rather than the original dataset and con- has many applications, from intrusion detection to infor-
verge faster. mation retrieval, disease diagnosis, and protein sequence
search. This review of the literature focuses on unsupervised
• Reduces the challenges associated with high- learning methodologies and algorithms and the numerous
dimensional data. Regression-based algorithms are assessment metrics used to evaluate the performance of un-
readily over-adaptable when dealing with high- supervised learning models. It also outlines the advantages
dimensional data. We avoid overfitting prediction and disadvantages of each study. This survey report will
algorithms by utilizing PCA to minimize the size of aid academics in determining which unsupervised learning
the training dataset in advance. algorithms or approaches to utilize for issue solving. Also,
2) Disadvantages which study field needs greater attention. The scope of
• The key components have low interpretability. Princi- this research is confined to commonly used unsupervised
pal components are linear combinations of the origi- learning techniques. Only research within the last five years
nal data’s features, but they’re not easy to understand. should be highlighted. We may operate more algorithms and
For example, it’s challenging to identify the dataset’s methodologies in the future to improve targeting.
most relevant properties after computing the main
components.
http:// journals.uob.edu.bh
 Int. J. Com. Dig. Sys. 13, No.1, 911-921 (Apr-23) 919

A. Acknowledgment [13] J. Gao, C. Zhong, X. Chen, H. Lin, and Z. Zhang, “Unsupervised

The authors would like to thank the referees for their learning for passive beamforming,” ieee communications letters,
vol. 24, no. 5, pp. 1052–1056, 2020.
careful reading and for their comments, which significantly
improved the paper. Additionally, thanks to Dr. Salman [14] M. Caron, P. Bojanowski, A. Joulin, and M. Douze, “Deep clustering
Qadri, (Associate Professor, Chairman Department of Com- for unsupervised learning of visual features,” in Proceedings of the
puter Science, MNS University of Agriculture, Multan, Pak- European conference on computer vision (ECCV), 2018, pp. 132–
istan) and Dr. Farrukh Jamal, (Assistant Professor, Depart- 149.
ment of Statistics, The Islamia University of Bahawalpur,
Pakistan) for his motivational support. [15] M. Akram, S. Siddique, and M. G. Alharbi, “Clustering algorithm
with strength of connectedness for m-polar fuzzy network models,”
B. Conflicts of Interest Mathematical Biosciences and Engineering, vol. 19, no. 1, pp. 420–
455, 2022.
The authors declare no conflicts of interest.
References [16] M. Z. Rodriguez, C. H. Comin, D. Casanova, O. M. Bruno,
D. R. Amancio, L. d. F. Costa, and F. A. Rodrigues, “Clustering
[1] J. Alzubi, A. Nayyar, and A. Kumar, “Machine learning from theory
algorithms: A comparative approach,” PloS one, vol. 14, no. 1, p.
to algorithms: an overview,” in Journal of physics: conference series,
e0210236, 2019.
vol. 1142, no. 1. IOP Publishing, 2018, p. 012012.
[17] S. Ahmadian, A. Epasto, R. Kumar, and M. Mahdian, “Cluster-
[2] Z. Zeng, Y. Li, Y. Li, and Y. Luo, “Statistical and machine learn-
ing without over-representation,” in Proceedings of the 25th ACM
ing methods for spatially resolved transcriptomics data analysis,”
SIGKDD International Conference on Knowledge Discovery & Data
Genome biology, vol. 23, no. 1, pp. 1–23, 2022.
Mining, 2019, pp. 267–275.
[3] J. G. Greener, S. M. Kandathil, L. Moffat, and D. T. Jones, “A guide
[18] A. Ali, W. K. Mashwani, M. H. Tahir, S. B. Belhaouari, H. Alraba-
to machine learning for biologists,” Nature Reviews Molecular Cell
iah, S. Naeem, J. A. Nasir, F. Jamal, and C. Chesneau, “Statis-
Biology, vol. 23, no. 1, pp. 40–55, 2022.
tical features analysis and discrimination of maize seeds utilizing
machine vision approach,” Journal of Intelligent & Fuzzy Systems,
[4] M. Alloghani, D. Al-Jumeily, J. Mustafina, A. Hussain, and A. J. Al- vol. 40, no. 1, pp. 703–714, 2021.
jaaf, “A systematic review on supervised and unsupervised machine
learning algorithms for data science,” Supervised and unsupervised
[19] C. S. Saba and N. Ngepah, “Convergence in renewable energy
learning for data science, pp. 3–21, 2020.
sources and the dynamics of their determinants: an insight from a
club clustering algorithm,” Energy Reports, vol. 8, pp. 3483–3506,
[5] R. Saravanan and P. Sujatha, “A state of art techniques on machine 2022.
learning algorithms: a perspective of supervised learning approaches
in data classification,” in 2018 Second International Conference on
[20] A. Maji, N. R. Velaga, and Y. Urie, “Hierarchical clustering analysis
Intelligent Computing and Control Systems (ICICCS). IEEE, 2018,
framework of mutually exclusive crash causation parameters for
pp. 945–949.
regional road safety strategies,” International journal of injury
control and safety promotion, vol. 25, no. 3, pp. 257–271, 2018.
[6] D. Krotov and J. J. Hopfield, “Unsupervised learning by competing
hidden units,” Proceedings of the National Academy of Sciences,
[21] H. Ma, “Design of chinese linguistics teaching system based on
vol. 116, no. 16, pp. 7723–7731, 2019.
k-means clustering algorithm,” in The International Conference on
Cyber Security Intelligence and Analytics. Springer, 2022, pp.
[7] H. U. Dike, Y. Zhou, K. K. Deveerasetty, and Q. Wu, “Unsupervised 424–431.
learning based on artificial neural network: A review,” in 2018 IEEE
International Conference on Cyborg and Bionic Systems (CBS).
[22] Y. Zhang, Y. Zhang, Q. Chen, Z. Ai, and Z. Gong, “True-link
IEEE, 2018, pp. 322–327.
clustering through signaling process and subcommunity merge in
overlapping community detection,” Neural Computing and Applica-
[8] N. Li, M. Shepperd, and Y. Guo, “A systematic review of unsuper- tions, vol. 30, no. 12, pp. 3613–3621, 2018.
vised learning techniques for software defect prediction,” Informa-
tion and Software Technology, vol. 122, p. 106287, 2020.
[23] M. Li, X. Zhu, and S. Gong, “Unsupervised person re-identification
by deep learning tracklet association,” in Proceedings of the Euro-
[9] E. Acar and B. Yener, “Unsupervised multiway data analysis: pean conference on computer vision (ECCV), 2018, pp. 737–753.
A literature survey,” IEEE transactions on knowledge and data
engineering, vol. 21, no. 1, pp. 6–20, 2008.
[24] M. V. Prasad, R. Balakrishnan et al., “Spatio-temporal association
rule based deep annotation-free clustering (star-dac) for unsuper-
[10] V. Kavitha and M. Punithavalli, “Clustering time series data stream- vised person re-identification,” Pattern Recognition, vol. 122, p.
a literature survey,” arXiv preprint arXiv:1005.4270, 2010. 108287, 2022.

[11] N. Grira, M. Crucianu, and N. Boujemaa, “Unsupervised and semi- [25] M. A. Kabir and X. Luo, “Unsupervised learning for network flow
supervised clustering: a brief survey,” A review of machine learning based anomaly detection in the era of deep learning,” in 2020 IEEE
techniques for processing multimedia content, vol. 1, pp. 9–16, 2004. Sixth International Conference on Big Data Computing Service and
Applications (BigDataService). IEEE, 2020, pp. 165–168.
[12] R. A. Bantan, A. Ali, S. Naeem, F. Jamal, M. Elgarhy, and C. Ches-
neau, “Discrimination of sunflower seeds using multispectral and [26] K. Kottmann, P. Huembeli, M. Lewenstein, and A. Acı́n, “Unsu-
texture dataset in combination with region selection and supervised pervised phase discovery with deep anomaly detection,” Physical
classification methods,” Chaos: An Interdisciplinary Journal of Review Letters, vol. 125, no. 17, p. 170603, 2020.
Nonlinear Science, vol. 30, no. 11, p. 113142, 2020.

http:// journals.uob.edu.bh
920 Samreen Naeem, et al.: An Unsupervised Machine Learning Algorithms: Comprehensive Review.

[27] H. Choi, M. Kim, G. Lee, and W. Kim, “Unsupervised learning ap- mean clustering technique,” International Journal of E-Health and
proach for network intrusion detection system using autoencoders,” Medical Communications (IJEHMC), vol. 10, no. 4, pp. 54–65,
The Journal of Supercomputing, vol. 75, no. 9, pp. 5597–5621, 2019. 2019.

[28] J.-H. Seong and D.-H. Seo, “Selective unsupervised learning-based [42] M. Mateen, J. Wen, S. Song, and Z. Huang, “Fundus image clas-
wi-fi fingerprint system using autoencoder and gan,” IEEE Internet sification using vgg-19 architecture with pca and svd,” Symmetry,
of Things Journal, vol. 7, no. 3, pp. 1898–1909, 2019. vol. 11, no. 1, p. 1, 2018.

[29] A. I. Károly, R. Fullér, and P. Galambos, “Unsupervised clustering [43] Y. Dong and S. J. Qin, “A novel dynamic pca algorithm for dynamic
for deep learning: A tutorial survey,” Acta Polytechnica Hungarica, data modeling and process monitoring,” Journal of Process Control,
vol. 15, no. 8, pp. 29–53, 2018. vol. 67, pp. 1–11, 2018.

[30] N. Urs, S. Behpour, A. Georgaras, and M. V. Albert, “Unsupervised [44] S. Ghani, S. Kumari, and A. Bardhan, “A novel liquefaction study
learning in images and audio to produce neural receptive fields: for fine-grained soil using pca-based hybrid soft computing models,”
a primer and accessible notebook,” Artificial Intelligence Review, Sādhanā, vol. 46, no. 3, pp. 1–17, 2021.
vol. 55, no. 1, pp. 111–128, 2022.
[45] D. H. Grossoehme, M. Brown, G. Richner, S. M. Zhou, and
[31] K. La Marca and H. Bedle, “Deepwater seismic facies and ar- S. Friebert, “A retrospective examination of home pca use and
chitectural element interpretation aided with unsupervised machine parental satisfaction with pediatric palliative care patients,” Ameri-
learning techniques: Taranaki basin, new zealand,” Marine and can Journal of Hospice and Palliative Medicine®, vol. 39, no. 3,
Petroleum Geology, vol. 136, p. 105427, 2022. pp. 295–307, 2022.

[32] P. Edastama, A. S. Bist, and A. Prambudi, “Implementation of data

mining on glasses sales using the apriori algorithm,” International
Journal of Cyber and IT Service Management, vol. 1, no. 2, pp.
159–172, 2021. SAMREEN NAEEM got her Bachelor’s
Degree in Information Technology (IT) from
[33] P. B. I. S. Putra, N. P. S. M. Suryani, and S. Aryani, “Analysis
of apriori algorithm on sales transactions to arrange placement of
Sargodha University Pakistan, after that she
goods on minimarket,” International Journal of Engineering and enrolled and completed her M.Phil Degree
Emerging Technology, vol. 3, no. 1, pp. 13–17, 2018. in Computer Science from The Islamia Uni-
versity of Bahawalpur, Pakistan. She is also
[34] X. Xie, G. Fu, Y. Xue, Z. Zhao, P. Chen, B. Lu, and S. Jiang, work as Lecturer Computer Science and IT
“Risk prediction and factors risk analysis based on ifoa-grnn and in reputed institutes in Pakistan. Now she is
apriori algorithms: Application of artificial intelligence in accident
prevention,” Process Safety and Environmental Protection, vol. 122,
pursuing a PhD at the Southeast University
pp. 169–184, 2019. China to complete her education.

[35] L. Jia, L. Xiang, and X. Liu, “An improved eclat algorithm based
on tissue-like p system with active membranes,” Processes, vol. 7,
no. 9, p. 555, 2019.
AQIB ALI got his Bachelor’s Degree in
[36] W. Mohamed and M. A. Abdel-Fattah, “A proposed hybrid algo- Computer (2017), after that he enrolled and
rithm for mining frequent patterns on spark,” International Journal completed his M.Phil. Degree in Computer
of Business Intelligence and Data Mining, vol. 20, no. 2, pp. 146–
Science (2020) from The Islamia University
169, 2022.
of Bahawalpur, Pakistan. He is also working
[37] C.-H. Chee, J. Jaafar, I. A. Aziz, M. H. Hasan, and W. Yeoh, “Al- as Lecturer Computer Science and IT in
gorithms for frequent itemset mining: a literature review,” Artificial reputed institutes in Pakistan. Now he is
Intelligence Review, vol. 52, no. 4, pp. 2603–2621, 2019. pursuing a Ph.D. degree from the Southeast
University China to complete his education.
[38] L. Gubu, D. Rosadi et al., “Robust mean–variance portfolio selection
using cluster analysis: A comparison between kamila and weighted
k-mean clustering,” Asian Economic and Financial Review, vol. 10,
no. 10, pp. 1169–1186, 2020.

[39] S. Park, Y. Hwang, and B.-J. Yang, “Unsupervised learning of

topological phase diagram using topological data analysis,” Physical Sania Anam got her Bachelor’s Degree
Review B, vol. 105, no. 19, p. 195115, 2022. in Computer (2007),Master degree in Com-
puter (2009) after that he enrolled and com-
[40] P. An, Z. Wang, and C. Zhang, “Ensemble unsupervised autoen- pleted his M.Phil. Degree in Computer Sci-
coders and gaussian mixture model for cyberattack detection,” ence (2016) from The Islamia University
Information Processing & Management, vol. 59, no. 2, p. 102844, of Bahawalpur, Pakistan. Since 2016, she is
2022.
working as Lecturer in Computer Science in
[41] M. Rashid, H. Singh, and V. Goyal, “Cloud storage privacy in
Govt Associate College for Women Ahmad
health care systems based on ip and geo-location validation using k- pur East ,Bahawalpur, Pakistan.

http:// journals.uob.edu.bh
 Int. J. Com. Dig. Sys. 13, No.1, 911-921 (Apr-23) 921

MUHAMMAD MUNAWAR AHMED department of Information Technology at

completed his Bachelor’s Degree in The Islamia University of Bahawalpur,
Computer (2005), after that he completed Pakistan. Currently, he is enrolled in Ph.D.
his MSCS Degree in session (2011-13) program at The Islamia University of
from The Islamia University of Bahawalpur, Bahawalpur, Pakistan.
Pakistan. He is also working as Lecturer,

http:// journals.uob.edu.bh

View publication stats