Handling Imbalanced Data through Re-sampling: Systematic Review

Abstract

Handling imbalanced data is an important issue that can affect the validity and reliability of the results. One common approach to addressing this issue is through re-sampling the data. Re-sampling is a technique that allows researchers to balance the class distribution of their dataset by either over-sampling the minority class or under-sampling the majority class. Over-sampling involves adding more copies of the minority class examples to the dataset in order to balance out the class distribution. On the other hand, under-sampling involves removing some of the majority class examples from the dataset in order to balance out the class distribution. It's also common to combine both techniques, usually called hybrid sampling. It is important to note that re-sampling techniques can have an impact on the model's performance, and it is essential to evaluate the model using different evaluation metrics and to consider other techniques such as cost-sensitive learning and anomaly detection. In addition, it is important to keep in mind that increasing the sample size is always a good idea to improve the performance of the model. In this systematic review, we aim to provide an overview of existing methods for re-sampling imbalanced data. We will focus on methods that have been proposed in the literature and evaluate their effectiveness through a thorough examination of experimental results. The goal of this review is to provide practitioners with a comprehensive understanding of the different re-sampling methods available, as well as their strengths and weaknesses, to help them make informed decisions when dealing with imbalanced data.

Full text

Indonesian Journal of Electrical Engineering and Informatics (IJEEI)
Vol. 11, No. 2, June 2023, pp. 503~514
ISSN: 2089-3272, DOI: 10.52549/ijeei.v11i2. 4471  503
Journal homepage: http://section.iaesonline.com/index.php/IJEEI/index
Handling Imbalanced Data through Re-sampling: Systematic
Review
Razan Yasir1, Abdelrahman Elsharif Karrar2, Waleed Ibrahim Osman3, Moez Mutasim4
1,3,4 College of Computer Studies, The National Ribat University, Sudan
2College of Computer Science and Engineering, Taibah University, Saudi Arabia
Article Info
ABSTRACT
Article history:
Received Jan 28, 2023
Revised Apr 29, 2023
Accepted Jun 17, 2023
Handling imbalanced data is an important issue that can affect the validity and
reliability of the results. One common approach to addressing this issue is
through re-sampling the data. Re-sampling is a technique that allows
researchers to balance the class distribution of their dataset by either over-
sampling the minority class or under-sampling the majority class. Over-
sampling involves adding more copies of the minority class examples to the
dataset in order to balance out the class distribution. On the other hand, under-
sampling involves removing some of the majority class examples from the
dataset in order to balance out the class distribution. It's also common to
combine both techniques, usually called hybrid sampling. It is important to
note that re-sampling techniques can have an impact on the model's
performance, and it is essential to evaluate the model using different evaluation
metrics and to consider other techniques such as cost-sensitive learning and
anomaly detection. In addition, it is important to keep in mind that increasing
the sample size is always a good idea to improve the performance of the model.
In this systematic review, we aim to provide an overview of existing methods
for re-sampling imbalanced data. We will focus on methods that have been
proposed in the literature and evaluate their effectiveness through a thorough
examination of experimental results. The goal of this review is to provide
practitioners with a comprehensive understanding of the different re-sampling
methods available, as well as their strengths and weaknesses, to help them
make informed decisions when dealing with imbalanced data.
Keyword:
Data Mining
Imbalance Data
Re-sampling
Over-sampling
Under-sampling
Hybrid Sampling
SMOTE
Copyright © 2023 Institute of Advanced Engineering and Science.
All rights reserved.
Corresponding Author:
Razan Yasir Eltayeb,
Department of Information Technology,
National Ribat University
Nile Street, buri, Khartoum, Sudan.
Email: razanyasir946@gmail.com
1. INTRODUCTION
Knowledge Base (KB) is field which refers to a specially designed resource for gathering and
processing knowledge in logical statement formats that define the relationship between graphical entities [1],
and the field of machine learning which refers to the capability of a machine to imitate intelligent human
behavior, those are the common fields in artificial Intelligence and both of them can have a class imbalanced
data problem. The class imbalance problem become greatest issue in data mining and the imbalanced data
appears in daily application [2] specially on the medical apps as [2] discussed and in academic apps like using
the citation analysis by main path analysis (MPA) [3], this means that most of the fields are challenged by the
imbalanced data problem, which comes from having too much or too little data in one class relative to another.
This might also lead to complications like over-fitting, where a model becomes highly dependent on the
training data and struggles to generalize new data. There are several strategies to deal with the imbalanced data
problem, including altering data collection process, preprocess the data and impute the missing values to

 ISSN: 2089-3272
IJEEI, Vol. 11, No. 2, June 2023: 503 – 514
504
achieve better results [4], use employing new algorithms that are less sensitive to imbalance, and over-sampling
or under-sampling data.
Numerous academics have made significant contributions in this field due to how challenging the
problem of imbalanced data. In this section, we highlight relevant contributions made by various research
studies. Various academics have investigated recurring issues, imbalanced class distributions in a variety of
professions, and assessments of earlier studies based on the benchmark datasets they’ve employed.
Additionally, they have discussed a comparison of various approaches and learning algorithms related to
imbalanced data distributions issue. As illustrated in Figure 1, the imbalanced data problem can manifest in a
variety of ways, such as having too much data in one class and too little in another.
Figure 1. Illustration of Imbalanced Data [5]
2. METHODS FOR IMBALANCE HANDLING AT THE DATA LEVEL
A dataset is definitely imbalanced when there is a significant imbalance in the class distribution, such
as a ratio of 1:100 or 1:1000 samples from the minority class to the majority class.This bias in the training
dataset could have an impact on many machine learning algorithms, while other algorithms could totally ignore
the minority class. This is significant, because the minority class generally relies more heavily on prediction.
2.1. Re-sampling techniques
In many real-world supervised learning problems, an available dataset is imbalanced because it
contains a significantly lower number of instances of one class than another. In general, it is necessary to
correctly predict the minor class, but due to the limited information available, this can be challenging. One
solution to this class imbalance issue is to re-sample the training dataset. The two main approaches for re-
sampling an imbalanced dataset are under-sampling, which means removing examples from the majority class,
and over-sampling, which requires duplicating examples from the minority class. However, there are several
re-sampling techniques, making it challenging to pick the best one. As depicted in Figure 2, the two main
approaches for re-sampling an imbalanced dataset are under-sampling and over-sampling.
Figure 2. Illustration of Over-sampling and Under-sampling [6]
2.1.1. Re-sampling techniques classification based
Under-sampling: When the volume of the majority class is lowered, information is lost and potentially
overbroad rules are generated. Despite its low computational cost, the under-sampling strategy is commonly
utilized alongside clustering algorithms to tackle this problem. K-means is a partitional approach that is well

IJEEI ISSN: 2089-3272 
Handling Imbalanced Data through Re-sampling: Systematic Review (Razan Yasir et al)
505
known and widely used, and it offers a general way to properly cluster data. K-means was employed in [7] to
tackle the problem of a regular clustering center that was susceptible to disruption and falls into the optimal
solution locally. In a separate article, K-means [8] has been used to introduce an alternate method of assigning
a value to a cluster, enhancing the quality of unbalanced data in the algorithm's outputs. The DBSCAN
algorithm is another well-known algorithm to balance the training set, this algorithm selects the majority class
samples that are most appropriate and completely ignores other majority class samples. It was utilized in [9] to
provide an information-theoretic-based unsupervised gene selection technique. The majority class examples
that are appropriate for the offered novel on [10] are picked, and unwanted majority class samples are
eliminated to balance the training set. All of the data points were divided into different levels based on their
local density, and on [11], the DBSCAN identified the lawmakers among all of the data points and constructed
initial clusters based on the identified representatives for further analysis of the clusters' properties. This led to
the suggestion of a unified clustering framework without imbalanced data. An application of the fuzzy principle
is made possible by the fuzzy algorithms used in the cluster field to determine sample values during the
undersampling phase. Furthermore, [12] provides many parallelization solutions for clustering algorithms for
both Intel and NVidia-based architectures by focusing on three frequently used fuzzy clustering approaches,
including FCM, GK-FCM, and FM. Lastly, based on a fuzzy concept, [13] discussed a clustering technique
that allows for simplicity in data balancing.
Some techniques, such as the nearest neighbor balanced dataset, concentrate on circumstances near
the decision boundary. that was adapted by [22] to present a quantum algorithm for data classification that
breaks up larger groups of data belonging to the same class into smaller groups with sub-labels to help create
boundaries between data with different labels and then creates a quantum circuit for classification that has
multiple control gates. The under-sampling method is combined with the C4.5 classification model in [23].
Over-sampling: To achieve a balanced class distribution, approaches that generate fictional data are
more flexible than those that modify the machine learning model. Any predictor can be used with class-
imbalanced datasets according to these over-sampling techniques, which also alter the training data. For this
purpose, a number of algorithms have been presented, although most of them are complex and frequently create
additional noise. The over-sampling algorithm K-Means is applied for class-imbalanced data. It helps grouping
by generating minority-class samples in appropriate and significant areas of the input space. The method avoids
generating noise while actually solving imbalances between and within classes. When compared to standard
K-means and other K-means-like algorithms, the cosine similarity measure in [14] is applied to assign data
points to clusters and enhance the accuracy for certain datasets. When the dataset is uneven, the K-cosine-
means algorithm fails. One further survey [15] mentioned a few evolutionary algorithms and dividing
calculations. According to the correlation of those algorithms, the primary expert of k-implies is less expensive
to evaluate, but a con is that it is more responsive to noisy data and outliers than fuzzy k-means and k-medoids.
According to the distribution of the particulate matter values, the K-means clustering algorithm in [16]
separates the area of the center of the fine dust distribution and then determines the coordinates of the optimal
point. With regard to the DBSCAN algorithm, [17] offers a new grid-based algorithm, GBCN, that uses input
parameters in clusters. The first parameter is determined automatically and computes the distance between
each element of a dataset and its k nearest neighbors. The second parameter sets the minimum number of
elements in clusters at five to provide high efficiency. A novel density-based clustering algorithm called
ADBSCAN, where "A" stands for "adaptive," was developed to balance the dataset [18]. This algorithm offers
a new way to identify the dataset's local high-density samples, and it can effectively adapt to datasets with
significant variations in density without using any user-inputted variables. By clustering the dataset, granting
each element a membership value, and finding the initial cluster center, the fuzzy C-means clustering algorithm
described in [19] reduces the time complexity. The goal of the researchers in [20] and [21] is to offer several
clustering algorithms and explain how they behave with imbalanced data, are frequently applied to particular
tasks, and are not general. Table 1 presents a summary of previous studies that have employed cluster-based
re-sampling techniques, the table includes information on the specific clustering method used, the dataset(s)
analyzed, and the main aim of the studies.
In an attempt to balance out the medical data in 2021, [24] paired the c4.5 algorithm with a cluster-
based over-sampling technique. High classification accuracy, quick computation times, and understandable
classification rules are all benefits of the C4.5 decision tree algorithm. An unique method for generating
artificial data points while keeping in mind the distribution of the data and the k nearest neighbors is the K
Nearest Neighbor Oversampling (KNNOR) algorithm. The KNNOR method has exceeded the most powerful
augmentation methods by enabling classifiers to obtain significantly higher accuracy after adding artificial
minority data points to imbalanced datasets. A method to enhance the performance of kNN on imbalanced data
has been proposed in [25] and has shown promise in its ability to handle imbalanced datasets without
compromising model size. This method uses GA to optimize both feature weights and class weights. On [26],
a different methodology is offered that makes use of CNN to automatically extract useful features of various

 ISSN: 2089-3272
IJEEI, Vol. 11, No. 2, June 2023: 503 – 514
506
fault classes from datasets with highly skewed data and then adopts data augmentation techniques to solve the
problem. Imbalanced data classification using the Random Forest Classification (RFC) technique has become
extremely popular in the modern application period since the beginning of 2020. Doing so enables [27] to use
random forest (RF) as a classifier for imbalanced data after attempting to reduce outliers using isolation forest
(iForest) and increasing the number of instances in the dataset in a balanced way. According to Table 2, many
previous studies have utilized re-sampling techniques in combination with classification-based approaches to
improve the performance of their algorithms.
Table 1. Re-sampling Techniques Cluster-Based
Author(s)
Aim of study
Imbalanc
ed data
Applied on
Methods
Rahmanian, M.
and Mansoori,
E.G. [9]
propose an unsupervised gene
selection scheme based on
information theoretic measures
yes
Bench mark
microarray gene
expression dataset
A similarity-based algorithm,
density-based clustering methods
Wang, Y. and
Wang, D. [11]
propose an effective variational
clustering algorithm called VDPC
No
20 benchmark
datasets
DBSCAN and DPC
Oh, Yoosoo and
Min, Seonghee.
[16]
performed the K-means clustering
algorithm to cluster feature datasets
No
air pollution stations
in the korean dataset
K-means
Kume, A. and
Walker, S.G. [8]
introduce an alternative way to
assign a value to a cluster
No
sequential manner
dataset
k–means
Starczewski, A.
and Scherer,
M.M. [15]
proposed a new grid-based
algorithm GBCN
Yes
several 2-
dimensional datasets
DBSCAN algorithm, new grid-
based algorithm
Rohan Mittal
[19]
explains the fuzzy C-means
clustering algorithm
No
image segmentation,
geo cover image
capture dataset
fuzzy C-means
Khan, M.K. and
Ahmed, S.M.
[14]
using the cosine similarity measure
to assign data points to clusters
Yes
homogeneous
datasets, namely the
Iris and Seeds
dataset
standard K-means and other K-
means-like algorithms
Cebrian, J.M.
and Imbernón,
B. [12]
introduces several parallelization
strategies for clustering algorithms
for both Intel and Nvidia-based
architectures
Yes
different types of
datasets
fuzzy clustering techniques such
as FCM, GK-FCM, and FM
Lukauskas, M.
and Ruzgas, T.
[20]
introduce different clustering
algorithms and present how these
clustering algorithms work
Yes
distinguishing
datasets
UNIC, k-Medoids (PAM),
Gaussian Mixture, TCLUST,
Trimmed k-means, Spectral
Clustering, Density-Based Spatial
Clustering, MULIC, DENCLUE,
SOMs (NeuralNet), SVM,
HIERDENC, deeply embedded
clustering, etc
Patibandla, R.L.
and
Veeranjaneyulu,
N. [15]
outlined a few dividing calculations
and Evolutionary Algorithms
Yes
different info
datasets
k-implies, k-medoids, Fuzzy k-
means, Expectation
Maximization, etc
Jian, S., Li, D.
and Yu, Y. [7]
proposed improved k-means
clustering algorithm
Yes
extensive data of
New York City taxis
improved K-means
Lukauskas,
Mantas &
Ruzgas, Tomas.
[22]
introduce different clustering
algorithms and present how these
clustering algorithms work
Yes
spherical in shape
datasets
UNIC, k-Medoids (PAM),
Gaussian Mixture, TCLUST,
Trimmed k-means, Spectral
Clustering, Density-Based Spatial
Clustering, MULIC, DENCLUE,
SOMs (NeuralNet), SVM,
HIERDENC, Deep embedded
clustering, etc.
Pratiwi, N.B.I.
and Saputro,
D.R.S. [13]
discussed the clustering algorithm
with an approximation based on a
fuzzy concept, which grants
simplicity in the process of
constructing clusters on datasets
No
non-linear dataset
fuzzy c-means and Fuzzy C-
Shells
Mirzaei, B.,
Nikpour, B. and
Nezamabadi-
Pour, H. [10]
presented a novel and effective
under-sampling technique is
presented to select the suitable
samples of the majority class
Yes
over fifteen
imbalanced datasets
DBSCAN algorithm
Li, H., Liu, X.,
Li, T. and Gan,
R. [18]
Propose a novel density-based
clustering algorithm. After using the
density estimator to filter noise
samples
No
several artificial and
real-world datasets
ADBSCAN algorithm

IJEEI ISSN: 2089-3272 
Handling Imbalanced Data through Re-sampling: Systematic Review (Razan Yasir et al)
507
Table 2. Re-sampling Techniques Classification-Based
Author(s)
Aim of study
Imbalanced
data
Applied on
Methods
Li, J. and Kais, S.
[22]
presented a quantum algorithm for
data classification based on the
nearest-neighbor learning algorithm
No
trained experimental
dataset
nearest-neighbor
Xu, Z. and Shen, D.
[24]
proposed a cluster-based
oversampling algorithm (KNSMOTE)
Yes
imbalanced medical
dataset
Synthetic minority
oversampling
technique (SMOTE)
and k-means algorithm
Ijaz, M.F., Attique,
M. and Son, Y. [27]
proposes a cervical cancer prediction
model (CCPM) that offers an early
prediction of cervical cancer using
risk factors as inputs
Yes
foremost prevailing
cancer in females
dataset
random forest (RF)
Saqlain, M., Abbas,
Q. and Lee, J.Y.
[26]
proposed a deep learning-based CNN-
WDI model to classify wafer map
defects in the semiconductor
fabrication process
No
real wafer map
dataset
CNN algorithm
Nugraha, W.,
Maulana, M.S. and
Sasongko, A. [23]
using the undersampling strategy with
clustering techniques and
classification model
Yes
ten imbalance
datasets from
KEEL-repository
undersampling and
c4.5
Shih, Y.H. and Ting,
C.K. [25]
proposed a method to improve the
performance of kNN by developing a
new algorithm on imbalanced datasets
and using it to optimize both feature
and class weights
Yes
up-sampling dataset
FCWkNN and KNN
algorithm
2.1.2. Synthetic Minority Over-sampling Technique (SMOTE)
One solution for imbalanced data is to over-sample the minority class. The simplest approach is to
replicate examples from the minority class, although these examples don't add any better knowledge to the
model. Instead, by combining the current examples, new ones can be created. For the minority class, data
augmentation techniques like SMOTE are used. To further illustrate this technique, refer to Figure 3 which
visually demonstrates the SMOTE process, and tell us how SMOTE generates new synthetic examples of the
minority class by interpolating between existing examples in the minority class.
Figure 3. SMOTE Illustration [28]
There are various studies, especially on SMOTE and extreme learning machine algorithms, including
[29] and [30]. The [31] develop a Majority Weighted Minority Over-sampling Technique (MOT), which
creates synthetic samples from the weighted information samples using the clustering method. Utilizing
SMOTE as a superior over-sampling technique with K-means to overcome the imbalanced learning problem
[32]. K-means SMOTE successfully addresses data imbalance while enhancing the quality of freshly produced
artificial data and reducing the production of noisy data. Table 3 presents a summary of previous studies that
have applied re-sampling techniques using SMOTE method, and it includes information about authors, aim of
study and the dataset used.

 ISSN: 2089-3272
IJEEI, Vol. 11, No. 2, June 2023: 503 – 514
508
Table 3. SMOTE
Author(s)
Aim of study
Imbalanced
data
Applied on
Technique
Methods
Goyal, A.,
Rathore, L.
and Kumar, S.
[29]
Using SMOTE to
ensemble learning
algorithms is a more
efficient classifier in
classifying imbalances
Yes
highly imbalanced
datasets
Classification
SMOTE method for
over-sampling
Tian, C., Zhou,
L., Zhang, S.,
and Zhao, Y.
[31]
present a new weighted
minority oversampling
technique for the
classification of
imbalanced datasets
Yes
generate synthetic
samples dataset
Classification
Weighted Minority
Oversampling
Technique
(MWMOTE)
Saah, D. and
Tenneson, K.
[32]
addressing both
(between and within)-
class, while effectively
overcoming data
imbalance
Yes
seven remote
sensing
benchmark
datasets
Clustering
K-means and
SMOTE algorithm
Last, F.,
Douzas, G. and
Bacao, F. [30]
presents a simple and
effective oversampling
method
Yes
90 benchmark
datasets
Clustering
K-means and
SMOTE algorithm
2.2. Hybrid-Sampling:
Hybrid-sampling is a technique that combines multiple over-sampling and/or under-sampling
methods to address class imbalance in a dataset. The idea is to use the strengths of different methods to
overcome the limitations of a single approach. Figure 4 illustrates how hybrid-sampling is used to combine
multiple over-sampling and under-sampling methods.
Figure 4. Illustration of hybrid-sampling [33]
Hybrid methods, on the other hand, include both under-sampling and over-sampling strategies. A
novel cluster ensemble approach (CES) was developed by [34], which generates a single partition in the
majority function and may cause compatibility problems with the cluster ensemble architecture's functionality.
In order to improve predictions in binary imbalanced classification issues by the end of 2020, the proposed
study [35] attempted to create a distribution-free super ensemble classifier, which is a hybridization of HDDT
and RBFN models.
Researchers in [36] uses ensemble classifiers and the K-means clustering algorithm to deal with
various under-sampling or over-sampling approaches. From a side perspective, researchers [37] provide a
method that based on the K-means clustering algorithm to help in the selection of the dataset's important
features, and the K-nearest neighbor (KNN) is then used to assess the classification performance in the
experiments. A year earlier, [38] optimized the K-means clustering algorithm on the Hadoop platform and
developed a first cluster center selection technique based on the basic ideas of the Hash algorithm.
DBSCN and K-mean cluster analysis are currently the two most popular and simple-to-use algorithms for
clustering and classification-based analysis of data [39] because they can handle the majority of cases

IJEEI ISSN: 2089-3272 
Handling Imbalanced Data through Re-sampling: Systematic Review (Razan Yasir et al)
509
effectively when the data has a lot of randomness and no obvious set to use as a parameter, as in the case of
linear or logistic regression algorithms.
Researchers in [40] and introduce a model to enhanced hybrid bag-boost model with a proposed re-
sampling technique. This model includes a proposed re-sampling method that uses edited nearest neighbor
(ENN) under-sampling to reduce noise and K-Means SMOTE as an over-sampling method.
By combining distance with a mean-based re-sampling strategy, random forest was employed by [41]
to suggest an over-sampling method and deploy synthetic samples for the minority class.
Table 4 provides a comprehensive overview of previous studies that have used the hybrid-sampling
technique, and it highlights the effectiveness of this method in addressing class imbalance and improving the
performance of classifiers.
Table 4. hybrid sampling
Author(s)
Aim of study
Imbalanced
data
Applied on
Technique
Methods
Bajal, E. And
Katara, V. [39]
Used easily implementable
algorithm for clustering
Yes
a renal
adenocarcinoma
dataset
Clustering
DBSCAN and K-
mean
Puri, A. and Kumar
Gupta, M. [40]
improved hybrid bag boost
with the proposed re-
sampling technique model
Yes
binary imbalanced
dataset
Clustering
K-Means
SMOTE
(Synthetic
Minority
Oversampling
Technique),
edited the nearest
neighbor (ENN)
Khan, T. and Tian,
W. [34]
proposed a new cluster
ensemble method (CES)
Yes
ten real-world
benchmark datasets
Clustering
cluster ensemble
method (CES)
GÜLDAL, S. [41]
reduce the imbalanced ratio
Yes
Different collected
datasets
Classification
Random Forest
(RF) and Support
Vector Machine
(SVM)
algorithms
Shahabadi, M.S.E.
and Tabrizchi, H.
[36]
propose a novel clustering-
based undersampling
method to create a balanced
dataset
Yes
44 benchmark
datasets from the
KEEL repository
Clustering
k-means
Chakraborty, T.
and Chakraborty,
A.K. [35]
proposed a novel
distribution-free super
ensemble classifier which is
a hybrid model for
improving predictions in
binary imbalanced
classification problems
Yes
small or medium-
sized datasets
Classification
hybridization of
HDDT and
RBFN models
Moslehi, F. and
Haeri, A. [37]
present a novel and
integrated technique to
select practical features in
the dataset and remove the
non-relevant features
No
non-relevant
features dataset
Clustering
K-means
Hou, X. [38]
optimized the K-means
clustering algorithm on the
Hadoop platform
No
Hadoop dataset
Clustering
K-means
algorithm
2.3. Combined Techniques for Computational Ensemble
The previous section discussed how to handle imbalanced data by re-sampling the original data to
create classes that are balanced. The combination of techniques that manage imbalanced datasets and improve
performance would be investigated in this section.

 ISSN: 2089-3272
IJEEI, Vol. 11, No. 2, June 2023: 503 – 514
510
The Markov Clustering (MCL2) approach was determined to be the best suitable one for the dataset
in order to reproduce local identical-by-descent (IBD) graphs. [42] applies the Mean Shift Clustering algorithm
in a framework for detecting malicious actions and attack patterns in cyberspace. According to an adaptive
threshold, the new overlap-based under-sampling (OBU) approaches were introduced on [43] to exclude
negative instances from the overlapping region. Over fifty of them are available right away through the widely
utilized Fundamental Clustering Problems Suite (FCPS) on [44]. By decreasing and balancing energy
consumption, the study showed energy-efficient clustering method presented by [45] sought to increase the
energy efficiency of WSNs. Individuals are grouped using the distributed clustering algorithm in [46] based
on their social relationships and COVID-19 risk levels for serious disease. Lastly, [47] assessed, using a
bagging approach, how the use of smaller pixels impacts lesion detection performance in general oncologic
PET imaging.
Table 5 is a useful resource for researchers looking to improve the performance of their classifiers by
using combined techniques for computational ensemble. It provides an overview of the various methods used
in previous studies and the results obtained from their application.
Table 5. combined methods
3. DISCUSSION
Re-sampling the training dataset is a common solution to address class imbalance in a dataset. There are three
main techniques for re-sampling an imbalanced dataset, which are under-sampling, over-sampling, and a
combination of both (hybrid approach). All the previous studies in this article have also used these techniques
as well as other methods, as illustrated in figure 5.
Author(s)
Aim of study
Imbalanced data
Applied on
Methods
Shemirani, R. and Belbin,
G.M. [48]
demonstrated the
shortcomings of LFR
the standard
benchmark algorithm,
in simulating local
identical-by-descent
(IBD) graphs
No
graphs dataset
Markov Clustering
algorithm
Thrun, M.C. and Stier, Q.
[44]
presented immediate
access to over fifty
fundamental clustering
algorithms
No
arbitrary sample
size dataset
CRAN in R, python
Insausti, X. and Zárraga-
Rodríguez, M. [46]
presented a distributed
clustering algorithm
that groups individuals
according to their
social contacts and the
risk level of severe
illness
No
severe illness
from the
COVID-19
dataset
distributed consensus
algorithm
Niranjana, R., Kumar, V.A.
and Sheen, S. [42]
monitoring framework
for deducing malicious
activities and attack
patterns in the
cyberspace
No
Darknet traffic
dataset
Mean Shift clustering
algorithm
Vuttipittayamongkol, P. and
Elyan, E. [43]
proposed new Overlap-
based Undersampling
(OBU) methods
Yes
well-known
imbalanced
datasets
Overlap-based
Undersampling
method
Lin, D. and Wang, Q. [45]
proposed a novel
energy-efficient
clustering algorithm
that aimed at
improving the energy
efficiency of WSNs via
reducing and balancing
energy consumption
Yes
energy
overhead
dataset
dual-cluster-head
mechanism
Morey, A.M., Noo, F. and
Kadrmas, D.J. [47]
evaluated how the use
of smaller pixels
affects lesion-detection
performance in general
oncologic PET
imaging
No
general PET
cancer imaging
dataset

IJEEI ISSN: 2089-3272 
Handling Imbalanced Data through Re-sampling: Systematic Review (Razan Yasir et al)
511
Figure 5. Complete papers review
From the cluster-based previous studies, results shows the analysis of cluster-based re-sampling
studies has two main advantages over the use of generalized estimating equations. Firstly, there is no need to
create a correlation structure as within-cluster re-sampling takes care of it implicitly. Secondly, generalized
estimating equations may not be effective when dealing with non-ignorable clusters, whereas within-cluster
based re-sampling is still appropriate. Additionally, the use of these techniques results in more accurate
predictions as they create homogeneous groups with similar prognostic characteristics and corresponding
survival patterns. Figure 6 provides an illustration of the cluster-based re-sampling techniques discussed in the
paper.
Figure 6. Cluster-Based Technique
Data imbalance can be addressed by adjusting the proportion of majority and minority classes in the
data set through techniques such as undersampling the majority class or oversampling the minority class. Many
researchers have attempted to address this issue through methods such as cost-sensitive classification and
resampling techniques. These methods aim to improve the accuracy of identifying minority data by comparing
the performance of different classification algorithms. As shown in figure 7, these techniques can effectively
distinguish minority data from majority data.
Figure 7. Classification Based Technique
37%
15%
10%
20%
18%
Previous Studies
Cluster
Classification
Smote
Hyprid
Combined

 ISSN: 2089-3272
IJEEI, Vol. 11, No. 2, June 2023: 503 – 514
512
Previous research suggests that using SMOTE (Synthetic Minority Over-sampling Technique) can
increase recall for minority class predictions, but at the cost of decreased precision. This means that more
minority class predictions will be made, but some may not be accurate. Studies have also found that SMOTE
can lead to more class overlap and noise due to not considering neighboring samples from different classes.
Additionally, SMOTE may not be as effective for data with high dimensions. However, it can reduce overfitting
and there is no information loss when using it. It is also easy to understand and manipulate. Figure 8 illustrates
the results of previous studies on SMOTE.
Figure 8. SMOTE Technique
Finally, figure 9 illustrate the hybrid approach. Taken together, all the studies presented previously in
this section suggest a hybrid sampling approach that combines two sampling techniques to create a balanced
dataset. Combining several sampling strategies enhances each method's strengths while reducing its drawbacks.
As a result, the hybrid-sampling technique regularly outperforms the individual sampling techniques.
Figure 9. Hybrid-sampling Technique
4. CONCLUSION
Re-sampling is the suitable way for handling imbalanced data, according to all of the prior studies in
this paper. The re-sampling and combined approaches can accomplish much more because they can provide
the learning algorithm with new information or get rid of unnecessary information. On small, imbalanced
datasets, the minority class appears to be poorly represented by an excessively reduced number of examples
that may not be sufficient for learning, especially when a high degree of class overlap exists and the class is
further categorized into subclusters. The above relationship between training set size and improper
classification or clustering performance for imbalanced datasets appears to be causal. The impact of these
complicated factors appears to be diminished for larger datasets as the minority class is better represented by a
greater number of instances.

IJEEI ISSN: 2089-3272 
Handling Imbalanced Data through Re-sampling: Systematic Review (Razan Yasir et al)
513
REFERENCES:
[1] A. E. Karrar, “A Proposed Model for Improving the Performance of Knowledge Bases in Real-World Applications
by Extracting Semantic Information,” International Journal of Advanced Computer Science and Applications, vol.
13, no. 2, 2022, doi: 10.14569/ijacsa.2022.0130214.
[2] A. E. Karrar, “Investigate the Ensemble Model by Intelligence Analysis to Improve the Accuracy of the Classification
Data in the Diagnostic and Treatment Interventions for Prostate Cancer,” International Journal of Advanced
Computer Science and Applications, vol. 13, no. 1, 2022, doi: 10.14569/ijacsa.2022.0130122.
[3] M. Umair et al., “Main Path Analysis to Filter Unbiased Literature,” Intelligent Automation & Soft Computing, vol.
32, no. 2, pp. 1179–1194, 2022, doi: 10.32604/iasc.2022.018952.
[4] A. E. Karrar, “The Effect of Using Data Pre-Processing by Imputations in Handling Missing Values,” Indonesian
Journal of Electrical Engineering and Informatics (IJEEI), vol. 10, no. 2, Apr. 2022, doi: 10.52549/ijeei.v10i2.3730.
[5] https://www.urbanstat.com/handlingimbalanceddatasets/?doing_wp_cron=1659547994.1132130622863769531250
[6] Xia, Wei & Ma, Caihong & Liu, Jianbo & Liu, Shibin & Chen, Fu & Zhi, Yang & Duan, Jianbo. (2019). High-
Resolution Remote Sensing Imagery Classification of Imbalanced Data Using Multistage Sampling Method and
Deep Neural Networks. Remote Sensing. 11. 2523. 10.3390/rs11212523.
[7] Jian, S., Li, D. and Yu, Y., 2021. ‘Research on Taxi Operation Characteristics by Improved DBSCAN Density
Clustering Algorithm and K-means Clustering Algorithm’. In Journal of Physics: Conference Series (Vol. 1952, No.
4, p. 042103). IOP Publishing.
[8] Kume, A. and Walker, S.G., 2021. ‘The utility of clusters and a Hungarian clustering algorithm’. Plos one, 16(8),
p.e0255174.
[9] Rahmanian, M. and Mansoori, E.G., 2022. ‘An unsupervised gene selection method based on multivariate normalized
mutual information of genes ’. Chemometrics and Intelligent Laboratory Systems, 222, p.104512.
[10] Mirzaei, B., Nikpour, B. and Nezamabadi-Pour, H., 2020. ‘An under-sampling technique for imbalanced data
classification based on DBSCAN algorithm’. In 2020 8th Iranian Joint Congress on Fuzzy and intelligent Systems
(CFIS) (pp. 21-26). IEEE.
[11] Wang, Y., Wang, D., Zhou, Y., Zhang, X. and Quek, C., 2021. ‘VDPC: Variational Density Peak Clustering
Algorithm’. arXiv preprint arXiv:2201.00641.
[12] Cebrian, J.M., Imbernón, B., Soto, J. and Cecilia, J.M., 2021. ‘Evaluation of Clustering Algorithms on HPC
Platforms’. Mathematics, 9(17), p.2156.
[13] Pratiwi, N.B.I. and Saputro, D.R.S., 2020, August. ‘Fuzzy c-shells clustering algorithm’. In Journal of Physics:
Conference Series (Vol. 1613, No. 1, p. 012006). IOP Publishing.
[14] M.K., Ahmed, S.M., Sarker, S. and Khan, M.H., 2021. ‘K-Cosine-Medoids Clustering Algorithm’. In 2021 5th
International Conference on Electrical Information and Communication Technology (EICT) (pp. 1-5). IEEE.
[15] Patibandla, R.L. and Veeranjaneyulu, N., 2021. ‘Clustering Algorithms: An Exploratory Review’.
[16] Oh, Yoosoo & Min, Seonghee. (2021). ‘Practical Application Using the Clustering Algorithm’. 10.5772/intechopen.
99314.
[17] Starczewski, A., Scherer, M.M., Ksiazek, W., Debski, M. and Wang, L., 2021. ‘A novel grid-based clustering
algorithm’. Journal of Artificial Intelligence and Soft Computing Research, 11.
[18] Li, H., Liu, X., Li, T. and Gan, R., 2020. ‘A novel density-based clustering algorithm using nearest neighbor graph’.
Pattern Recognition, 102, p.107206.
[19] Mittal, Rohan. (2021). ‘Fuzzy C-Means Clustering Algorithm’.
[20] Lukauskas, M. and Ruzgas, T., 2021. ‘Analysis of clustering methods performance across multiple datasets’. In
DAMSS 2021: 12th conference on data analysis methods for software systems, Druskininkai, Lithuania, December
2–4, 2021 (pp. 45-46). Vilnius university press.
[21] Lukauskas, Mantas & Ruzgas, Tomas. (2021). ‘Comparative analysis of clustering algorithms for synthetic and real
data’.
[22] Li, J. and Kais, S., 2021. ‘A universal quantum circuit design for periodical functions’. New Journal of Physics,
23(10), p.103022.
[23] Nugraha, W., Maulana, M.S. and Sasongko, A., 2020. ‘Clustering Based Undersampling for Handling Class
Imbalance in C4.5 Classification Algorithm’. In Journal of Physics: Conference Series (Vol. 1641, No. 1, p. 012014).
IOP Publishing.
[24] Xu, Z., Shen, D., Nie, T., Kou, Y., Yin, N. and Han, X., 2021. ‘A cluster-based oversampling algorithm combining
SMOTE and k-means for imbalanced medical data’. Information Sciences, 572, pp.574-589.
[25] Shih, Y.H. and Ting, C.K., 2019. ‘Evolutionary optimization on k-nearest neighbors classifier for imbalanced
datasets’. In 2019 IEEE Congress on Evolutionary Computation (CEC) (pp. 3348-3355). IEEE.
[26] Saqlain, M., Abbas, Q. and Lee, J.Y., 2020. ‘A deep convolutional neural network for wafer defect identification on
an imbalanced dataset in semiconductor manufacturing processes’. IEEE Transactions on Semiconductor
Manufacturing, 33(3), pp.436-444.

 ISSN: 2089-3272
IJEEI, Vol. 11, No. 2, June 2023: 503 – 514
514
[27] Ijaz, M.F., Attique, M. and Son, Y., 2020. ‘Data-driven cervical cancer prediction model with outlier detection and
over-sampling methods’. Sensors, 20(10), p.2809.
[28] Vijayvargiya, Ankit & Prakash, Chandra & Kumar, Rajesh & Bansal, Sanjeev & Tavares, Joao. (2021). Human Knee
Abnormality Detection from Imbalanced sEMG Data. Biomedical Signal Processing and Control. 66.
10.1016/j.bspc.2021.102406
[29] Goyal, A., Rathore, L. and Kumar, S., 2021. ‘A Survey on Solution of Imbalanced Data Classification Problem Using
SMOTE and Extreme Learning Machine’. In Communication and Intelligent Systems (pp. 31-44). Springer,
Singapore.
[30] Last, F., Douzas, G. and Bacao, F., 2017. ‘Oversampling for imbalanced learning based on k-means and smote’.
arXiv preprint arXiv:1711.00837.
[31] Tian, C., Zhou, L., Zhang, S. and Zhao, Y., 2020. ‘A new majority weighted minority oversampling technique for
classification of imbalanced datasets’. In 2020 International Conference on Big Data, Artificial Intelligence and
Internet of Things Engineering (ICBAIE) (pp. 154-157). IEEE.
[32] Saah, D., Tenneson, K., Matin, M., Uddin, K., Cutter, P., Poortinga, A., Nguyen, Q.H., Patterson, M., Johnson, G.,
Markert, K. and Flores, A., 2019. ‘Land cover mapping in data scarce environments: challenges and opportunities’.
Frontiers in Environmental Science, 7, p.150.
[33] Tuong Le, Minh Thanh Vo, Bay Vo, Mi Young Lee, Sung Wook Baik, "A Hybrid Approach Using Oversampling
Technique and Cost-Sensitive Learning for Bankruptcy Prediction", Complexity, vol. 2019, Article ID 8460934, 12
pages, 2019. https://doi.org/10.1155/2019/8460934
[34] Khan, T., Tian, W., Kadhim, M.R. and Buyya, R., 2021. ‘A Novel Cluster Ensemble based on a Single Clustering
Algorithm’. In 2021 16th Conference on Computer Science and Intelligence Systems (FedCSIS) (pp. 127-135). IEEE.
[35] Chakraborty, T. and Chakraborty, A.K., 2020. ‘Superensemble classifier for improving predictions in imbalanced
datasets’. Communications in Statistics: Case Studies, Data Analysis and Applications, 6(2), pp.123-141.
[36] Shahabadi, M.S.E., Tabrizchi, H., Rafsanjani, M.K., Gupta, B.B. and Palmieri, F., 2021. ‘A combination of
clustering-based under-sampling with ensemble methods for solving imbalanced class problem in intelligent
systems’. Technological Forecasting and Social Change, 169, p.120796.
[37] Moslehi, F. and Haeri, A., 2020. ‘An evolutionary computation-based approach for feature selection’. Journal of
Ambient Intelligence and Humanized Computing, 11(9), pp.3757-3769.
[38] Hou, X., 2019. ‘An improved k-means clustering algorithm based on hadoop platform’. In The International
Conference on Cyber Security Intelligence and Analytics (pp. 1101-1109). Springer, Cham.
[39] Bajal, E., Katara, V., Bhatia, M. and Hooda, M., 2022. ‘A Review of Clustering Algorithms: Comparison of
DBSCAN and K-mean with Oversampling and t-SNE’. Recent Patents on Engineering, 16(2), pp.17-31.
[40] Puri, A. and Kumar Gupta, M., 2022. ‘Improved hybrid bag-boost ensemble with K-means-SMOTE–ENN technique
for handling noisy class imbalanced data.’ The Computer Journal, 65(1), pp.124-138.
[41] GÜLDAL, S., 2021. ‘Improving Machine Learning Performance of Imbalanced Data by Resampling: DBSCAN and
Weighted Arithmetic Mean’. Bitlis Eren Üniversitesi Fen Bilimleri Dergisi, 10(4), pp.1563-1574.
[42] Niranjana, R., Kumar, V.A. and Sheen, S., 2020. ‘Darknet traffic analysis and classification using numerical agm and
mean shift clustering algorithm’. SN Computer Science, 1(1), pp.1-10.
[43] Vuttipittayamongkol, P. and Elyan, E., 2020. ‘Neighbourhood-based undersampling approach for handling
imbalanced and overlapped data’. Information Sciences, 509, pp.47-70.
[44] Thrun, M.C. and Stier, Q., 2021. ‘Fundamental clustering algorithms suite’. SoftwareX, 13, p.100642.
[45] Lin, D. and Wang, Q., 2019. ‘An energy-efficient clustering algorithm combined game theory and dual-cluster-head
mechanism for WSNs’. IEEE Access, 7, pp.49894-49905.
[46] Insausti, X., Zárraga-Rodríguez, M., Nolasco-Ferencikova, C. and Gutierrez-Gutierrez, J., 2021. ‘Distributed
clustering algorithm for adaptive pandemic control’. IEEE Access, 9, pp.160688-160696.
[47] Morey, A.M., Noo, F. and Kadrmas, D.J., 2016. ‘Effect of using 2 mm voxels on observer performance for PET
lesion detection’. IEEE transactions on nuclear science, 63(3), pp.1359-1366.
[48] Shemirani, R., Belbin, G.M., Burghardt, K., Lerman, K., Avery, C.L., Kenny, E.E., Gignoux, C.R. and Ambite, J.,
2021. ‘Selecting Clustering Algorithms for IBD Mapping’. bioRxiv.