www.nature.

com/scientificreports

OPEN Hybrid clustering strategies

for effective oversampling
and undersampling in multiclass
classification
Amirreza Salehi 1 & Majid Khedmati 2*
Multiclass imbalance is a challenging problem in real-world datasets, where certain classes may have
a low number of samples because they correspond to rare occurrences. To address the challenge
of multiclass imbalance, this paper introduces a novel hybrid cluster-based oversampling and
undersampling (HCBOU) technique. By clustering and separating classes into majority and minority
categories, this algorithm retains the most information during undersampling while generating
efficient data in the minority class. The classification is carried out using one-vs-one and one-vs-all
decomposition schemes. Extensive experimentation was carried out on 30 datasets to evaluate
the proposed algorithm’s performance. The results were subsequently compared with those of
several state-of-the-art algorithms. Based on the results, the proposed algorithm outperforms
the competing algorithms under different scenarios. Finally, The HCBOU algorithm demonstrated
robust performance across varying class imbalance levels, highlighting its effectiveness in handling
imbalanced datasets.

Keywords Multiclass classification, Imbalanced data, Oversampling, Undersampling, Ensemble

Classification is one of the data mining techniques used to predict the class label of instances based on their
features. Although most research efforts concentrate on two-class classification, multiclass classification is one
of the most challenging machine learning research t­ opics1. There exist widespread applications of multiclass
classification in machine learning where fraud d ­ etection2, disease ­diagnosis3, Sentiment ­analysis4, plant species
­recognition5, and image c­ lassification6are some examples of this problem. On the other hand, it is important to
acknowledge that numerous real-world datasets exhibit a significant disparity in the number of instances across
different ­classes7. This is known as the multiclass imbalance classification problem. Given this problem, the
machine learning models may perform poorly for the minority class due to bias toward the majority c­ lass8. To
deal with the imbalance issue in classification, several methods have been proposed in the literature, which can
be classified into four categories. Algorithm-level approaches mitigate class imbalance by implementing modi-
fications and enhancements to classification algorithms. Data-level approaches address the class imbalance by
adjusting the dataset through techniques such as undersampling the majority class or oversampling the minority
class to mitigate the effects of skewed class distribution on the learning process. Cost-sensitive approaches seek to
minimize the total misclassification cost by integrating algorithm-level and data-level strategies. Finally, ensem-
ble methods, such as bagging and boosting, combine multiple classifiers to improve predictive performance.
However, in most cases, the ensemble algorithms are combined with data-level or algorithm-level ­techniques9.
Multiclass classification, involving more than two classes, presents significant challenges due to the increased
complexity of decision boundaries. In contrast, building a classifier to distinguish only between two classes can
be easier since the decision boundaries can be more straightforward. Multiclass classification problems can be
addressed by decomposing them into binary subproblems using various techniques, such as one-versus-one
(OVO) and one-versus-all (OVA). The OVO technique divides a multiclass problem into many binary problems,
with each binary classifier learning to differentiate between a pair of classes. The final predicted class is deter-
mined by aggregating the outputs of the individual base classifiers. The OVA technique trains a binary classifier

1
Department of Industrial Engineering, Sharif University of Technology, Tehran, Iran. 2Department of Industrial
Engineering, Sharif University of Technology, Azadi Ave., Tehran 1458889694, Iran. *email: [email protected]

Scientific Reports | (2025) 15:3460 | https://doi.org/10.1038/s41598-024-84786-2 1

Vol.:(0123456789)
 www.nature.com/scientificreports/

for each class to differentiate it from all other classes. When a base classifier yields a positive prediction, the cor-
responding class is assigned as the ­output10. Despite the existence of multiple methods proposed in the literature
for addressing imbalanced data, there exist some drawbacks to these methods. In this regard,

• Existing balancing methods primarily address binary datasets, and fewer approaches have been proposed to
handle multiclass decomposition schemes in real-world datasets.
• The utilization of oversampling techniques leads to an expansion in the dataset size, which poses challenges
when applying learning operations to the dataset.
• Undersampling techniques, on the other hand, result in a significant loss of information due to the removal
of samples.
• Oversampling techniques that involve duplicating samples can lead to overfitting on the training dataset,
compromising the generalization capabilities of the model.
• The random nature of undersampling and oversampling techniques may not accurately represent the true
features of the dataset, causing potential biases in data.
• Algorithm-level and ensemble algorithms, when applied independently, often exhibit inconsistent perfor-
mance and are susceptible to the challenges posed by imbalanced datasets.
• Sampling from the majority classes in highly imbalanced datasets can result in model underfitting on majority
classes, affecting the overall performance.
• Methods based on clustering often disregard the small number of samples present in each cluster, which can
affect the effectiveness of these methods.
• Some methods for multiclass classification cannot be applied in different decomposition schemes, limiting
their usability.

In this paper, we propose a novel hybrid cluster-based oversampling and undersampling (HCBOU) algorithm
for classifying multiclass imbalanced datasets that combines two data-level techniques, including oversampling
and undersampling, with K-means clustering. The proposed algorithm utilizes both OVO and OVA decompo-
sition techniques for classification. The combined application of oversampling and undersampling techniques
can effectively mitigate the risks of overfitting, which may arise from exclusive reliance on oversampling, and
information loss, which can occur when undersampling is applied excessively. In this approach, the classes of
datasets are categorized into two groups, including majority and minority, depending on the number of data
instances. This categorization would result in multiple majority or minority classes. Then, the algorithm employs
a clustering technique within each minority group to identify relevant clusters and generate more meaningful
data. Furthermore, sampling is carried out in each majority class using a clustering-based approach, which
minimizes the influence of noisy data and effectively reduces information loss. The objectives and contributions
of this research study are summarized as follows:

• We propose a novel hybrid approach that combines undersampling, oversampling, and clustering techniques
to address class imbalance. This algorithm effectively balances multiclass imbalanced datasets by combin-
ing the strengths of both data-level methods and ensemble learning. By leveraging K-means clustering, we
ensure more meaningful data generation for minority classes while minimizing noise and information loss
in majority classes.
• The effectiveness and reliability of HCBOU are thoroughly validated through extensive experiments con-
ducted on 30 datasets, each with varying degrees of class imbalance. The proposed algorithm demonstrates
superior performance compared to six state-of-the-art algorithms across multiple evaluation metrics, includ-
ing precision, recall, and F1 score. This comprehensive assessment underlines the generalizability and consist-
ency of HCBOU across a wide range of real-world imbalanced datasets.
• The proposed HCBOU algorithm employs both OVO and OVA decomposition schemes, ensuring its flex-
ibility and robustness across different multiclass scenarios. This dual decomposition not only addresses the
complexity of multiclass imbalanced classification but also enables more precise decision boundaries by
simplifying complex multiclass problems into manageable binary sub-problems.

The remainder of the paper is organized as follows. Section “Literature review” provides a comprehensive
review of existing research on classifying imbalanced multi-class datasets. The details of the proposed approach
are provided in Sect.“The proposed HCBOU algorithm”. Section “Experiments” provides the experimental
setup. The performance of the proposed approach is evaluated and compared to the competing algorithms in
Sect. “Results”. Finally, Sect. :“Conclusions”. provides the conclusions and some future research recommendations.

Literature review
As mentioned in the previous section, data balancing methods can be classified into four categories. In this sec-
tion, a comprehensive literature review of research studies on data balancing methods is presented.

Data‑level methods
Data-level approaches address the class imbalance by adjusting the training dataset through techniques such
as oversampling or undersampling. Oversampling involves producing more instances of the minority class by
duplicating existing instances or generating new synthetic instances using methods like Synthetic Minority Over-
sampling Technique (SMOTE)11. On the other hand, undersampling entails eliminating a subset of instances from
the majority class using methods like Tomek ­links12 or randomly deleting instances from the majority class. To

Scientific Reports | (2025) 15:3460 | https://doi.org/10.1038/s41598-024-84786-2 2

Vol:.(1234567890)
www.nature.com/scientificreports/

name a few research works in the context of data-level methods, Gao et al.13 proposed the differential partition
sampling ensemble method (DPSE) within the OVA framework. They classified all samples into safe examples,
borderline examples, rare examples, and outliers. According to the distributional characteristics of the classes,
random undersampling for safe examples and SMOTE for borderline and rare examples are then offered. Finally,
they concluded that the proposed DPSE performs better than traditional techniques in OVA, OVO, and direct
classification schemes. Krawczyk et al.14 introduced multiclass radial-based oversampling (MCRBO), a technique
that employs potential functions to synthesize new instances. Finding regions with low mutual class distribu-
tion values guides synthetic instance generation. Li et al.15 presented a clustering-based technique for multiclass
imbalanced problems. They initially split a multiclass dataset into several binary-class datasets and then, using
spectral clustering divided the minority classes of the binary-class subsets into subspaces. After subspace iden-
tification, oversampling is performed tailored to the characteristics of each subspace. Liu et al.16 proposed a real-
value negative selection detector-based oversampling method that modifies the traditional real-value negative
selection technique to handle the multiclass imbalance problem. The loss of information of minority classes is
reduced by minimizing the within-class imbalance. Neetha et al.17 proposed Borderline-DEMNET, addressing
the class imbalance in Alzheimer’s disease classification using Borderline-SMOTE, achieving high accuracy and
outperforming previous multiclass classification models. Arafat et al.18 proposed a cluster-based undersampling
approach to handle imbalanced multiclass classification problems. The suggested approach divides instances of
the majority class into clusters, chooses the most useful instances within each cluster to generate several bal-
anced datasets, and finally applies the random forest algorithm to the balanced datasets. Dai et al.19 proposed
the Three-line Hybrid Positive Instance Augmentation (THPIA) algorithm, leveraging genetic principles to mix
features of majority and minority classes, improving minority instance representation and reducing overfitting.

Algorithm‑level methods
The algorithm-level approach entails explicitly addressing the issue of class imbalance by applying some improve-
ments in the existing learning algorithms or developing novel algorithms. In this regard, Chen et al.20proposed a
double kernel-based class-specific broad learning system (DKCSBLS) to cope with multiclass imbalanced data.
To put more emphasis on minority classes, the model includes class-specific penalty coefficients. Furthermore, a
double kernel mapping approach aims to capture features with increased robustness. Vij and ­Arora21 proposed a
deep transfer learning-based diagnostic system for multiclass diabetic retinopathy classification, using modified
models (Inception V3, ResNet34, EfficientNet B0, VGG16, Xception) to enhance diagnosis accuracy, achiev-
ing 99.36% accuracy with balanced imbalanced data labels. Ding et al.22proposed a weighted online sequential
extreme learning machine with kernels (WOS-ELMK) for both binary class and multiclass imbalance learning.
The non-optimal hidden node problem related to random feature mapping in previous online sequential extreme
learning machine (OS-ELM) approaches for imbalance learning is avoided by their proposed kernel mapping in
WOS-ELMK. Ketu and M ­ ishra23 proposed a scalable kernel-based SVM classification approach, which is based
on the concept of the adjusting kernel scaling (AKS) approach to handle the multiclass imbalanced dataset. The
chi-square test and weighting criteria have been used to evaluate the selection of the kernel function. Li et al.24
proposed a multiclass imbalance classification approach that incorporates a class imbalance ratio, density-based
factor, and adaptive weighting mechanism, enhancing the distribution of weights for better handling imbalance
issues. Dai et al.25 introduced a novel Schur decomposition class-overlap undersampling method (SDCU) to
globally identify overlapping instances, enhancing classifier performance in imbalanced datasets by reducing
overlap and noise. Han et al.26 proposed a global–local-based oversampling method (GLOS) that adjusts synthetic
instance generation based on class-level and instance-level dispersion, improving the quality and relevance of
the synthetic data for multiclass imbalance problems.

Cost‑sensitive methods
Cost-sensitive learning explicitly accounts for the misclassification costs while training a model and minimizes
the expected cost of misclassification. For an imbalanced dataset, misclassifying the minority class is costlier than
misclassifying the majority classes. Tapkan et al.27 presented a cost-sensitive approach that utilizes the Bees algo-
rithm. The most significant advantage of this approach is its ability to handle binary and multiclass classification
problems. Additionally, it can incorporate misclassification costs into the algorithm by generating neighboring
solutions and evaluating the quality of the outcomes. Fernández et al.28 introduced a multiclass cost-sensitive
classification technique called Boosting Adapted for Cost matrix (BAdaCost). It involves combining several cost-
sensitive multiclass weak learners to create a powerful classification rule within the Boosting framework. Iran-
mehr et al.29 provided a constructive approach to improve the classifier about class imbalance by extending the
basic SVM loss function. It can be demonstrated that the resulting classifier ensures Bayes consistency. Liu et al.30
developed a multiclass imbalanced and concept drift network traffic classification framework based on online
active learning (MicFoal), addressing multiclass imbalance and concept drift in network traffic classification with
online active learning, showing superior performance on eight real-world datasets. Yang et al.31introduced a deep
reinforcement learning framework for handling multiclass imbalanced data in healthcare, enhancing minority
class prediction by combining dueling and double deep Q-learning with custom reward functions. Mienye and
­Sun32 proposed robust cost-sensitive classifiers that efficiently predict medical diagnosis by modifying the objec-
tive functions of well-known algorithms, including logistic regression, decision tree, extreme gradient boosting,
and random forest. Subsequently, the corresponding cost-sensitive algorithms for these models are developed.
Unlike resampling techniques, this approach does not modify the original data distribution.

Scientific Reports | (2025) 15:3460 | https://doi.org/10.1038/s41598-024-84786-2 3

Vol.:(0123456789)
 www.nature.com/scientificreports/

Ensemble methods
As mentioned previously, the ensemble approaches apply multiple classifiers to improve learning accuracy. Com-
bining ensemble algorithms with data-level approaches, the final performance of the model can be improved
during the learning process. Liu et al.33 proposed an EasyEnsemble approach that deals with imbalanced datasets.
In this approach, they created several subsets from the majority class where each subset is used to train a learner,
and their outputs are combined to form a final prediction. Seiffert et al.34 proposed the RUSBoost algorithm,
which combines undersampling and boosting approaches. Specifically, it utilizes random undersampling (RUS) to
remove instances randomly from the majority class until the desired balance is achieved. Grina et al.35 proposed
a re-sampling method based on belief function theory and ensemble learning, which assigns soft evidential
labels to improve object selection for both undersampling and oversampling in multiclass imbalance. Wang
et al.36 proposed an algorithm entitled SMOTEBagging that investigates the effects of diversity on each class. The
model combines SMOTE with bagging to handle imbalanced classification problems. Hido et al.37 presented the
Roughly Balanced Bagging method, which employs a novel sampling technique to enhance the original bagging
algorithm for imbalanced datasets. Wang et al.38 presented a technique entitled AdaBoost.NC, as a combination
of the multiclass classification AdaBoost algorithm with negative correlation learning. The initial weights given
to the training examples in this approach are inversely correlated with the number of instances in each class.
This approach enables the algorithm to better capture the complex relationships between different classes. Zhang
et al.39 proposed an efficient framework entitled SMOTE + AdaBoost. To balance the dataset before AdaBoost, the
SMOTE technique is used to generate synthetic majority classes. Chen et al.40 proposed a Balanced Random For-
est that reduces bias toward the majority class and improves the accuracy of predictions on the minority class by
randomly undersampling the majority class during the creation of each decision tree in the ensemble. Rodriguez
et al.41 proposed the Random Balance strategy (RandBal) for creating classifier ensembles to deal with imbalanced
two-class datasets. In RandBal, each base classifier is trained on a subset of data with a randomly assigned class
distribution, regardless of an apriori distribution. Consequently, for each subset, one class will be undersampled
while the other will be oversampled. Two approaches are available for handling multiclass problems: the first
approach decomposes the problem into a series of binary problems, while the second approach, Multiple Random
Balance (MultiRandBal), addresses all classes simultaneously. Dai et al.42 proposed a Heterogeneous Clustering
Ensemble learning method for Multiple Class-overlap Detection (HCE-MCD) to address multiclass imbalance
problems. The method uses a genetic algorithm to select and combine heterogeneous clustering techniques for
effective overlap detection and utilizes majority voting for improved clustering results.
The proposed HCBOU algorithm distinguishes itself from existing methods by integrating clustering-based
oversampling and undersampling techniques in a hybrid framework, ensuring more informed instance selection
and generation. Unlike traditional oversampling or undersampling approaches, HCBOU leverages K-means
clustering to improve data balance while minimizing noise and information loss. Additionally, HCBOU’s dual
decomposition using both OVO and OVA schemes ensures enhanced classification performance in multiclass
imbalanced datasets, offering a more flexible and robust solution compared to the algorithm-level, cost-sensitive,
and ensemble methods reviewed in the literature. The properties of the approaches stated in the literature review
are summarized in Table 1.

The proposed HCBOU algorithm

Despite the advancements in data-level, algorithm-level, and cost-sensitive approaches to handling class imbal-
ance, existing methods still face significant limitations in effectively managing multiclass imbalanced datasets.
Common oversampling techniques like SMOTE often lead to overfitting, while undersampling methods may
cause loss of crucial information. Ensemble-based methods can mitigate these issues but often require extensive
computational resources and are not always adaptable to complex class distributions. To address these challenges,
we propose the HCBOU algorithm. HCBOU uniquely integrates clustering techniques to inform both oversam-
pling and undersampling strategies, providing a more balanced and noise-free dataset while preserving minority
class representation. This framework is specifically motivated by the need for improved classification performance
and robustness in highly imbalanced, multiclass scenarios, where existing techniques either overgeneralize or
introduce bias. In this section, we propose a novel algorithm called HCBOU for classifying multiclass imbalanced
data. The details of the HCBOU algorithm are provided in the next subsection.

HCBOU: Hybrid cluster‑based oversampling and undersampling

Multiclass imbalance poses a significant challenge in real-world datasets. Oversampling and undersampling
methods have been proposed to tackle class imbalance, but a combination of these approaches can minimize
information loss and overfitting risks. Despite the progress made in this area, achieving a balanced representa-
tion of all classes while maintaining data quality remains an ongoing issue, particularly when complex multiclass
interactions are present. In this regard, maintaining the number of instances during data balancing is crucial
for overall model simplicity. Furthermore, the impact of overlapping data points, which often leads to ambigu-
ous decision boundaries, needs to be handled delicately to enhance model robustness. However, two key issues
persist that need to be addressed: determining the data for new instance generation and selecting data from the
majority class. This paper proposes a novel approach that combines oversampling, undersampling, and cluster-
ing methods to optimize sampling and data generation quality. Moreover, the algorithm integrates both OVO
and OVA decomposition approaches to evaluate the performance across different decomposition schemes. This
dual decomposition strategy enables a more granular evaluation of the model’s performance across various
classes, providing insights into potential weaknesses in classification accuracy for minority classes. The proposed
algorithm, illustrated in Fig. 1, demonstrates promising results in addressing the multiclass imbalance problem.
The notations used in this paper are represented in Table 2.

Scientific Reports | (2025) 15:3460 | https://doi.org/10.1038/s41598-024-84786-2 4

Vol:.(1234567890)
www.nature.com/scientificreports/

Data level Ensemble Scheme

Algorithm Undersampling Oversampling Cost-sensitive Algorithm level Bagging Boosting OVO OVA Clustering
Differential
partition sampling ✔ ✔ ✔ ✔ ✔
ensemble (DPSE)
Multiclass radial-
based oversam- ✔ ✔ ✔
pling (MCRBO)
OVA decomposi-
tion and spectral ✔ ✔ ✔
clustering
Modified real-value
negative selection ✔ ✔
oversampling
Cluster-based
✔ ✔ ✔
Undersampling
Double-kernel
based class-specific
✔
broad learning
system
Kernel based
✔
online learning
BEE-Miner ✔
BAdaCost ✔ ✔
Cost-sensitive
support vector ✔
machines
EasyEnsemble ✔ ✔ ✔
RUSBoost ✔ ✔ ✔
SMOTEBagging ✔ ✔ ✔
Roughly Balanced
✔ ✔ ✔
Bagging
AdaBoost.NC ✔ ✔ ✔ ✔ ✔
SMOTE + Ada-
✔ ✔ ✔
Boost
Balanced Random
✔ ✔
Forest
Scalable
kernel‑based SVM ✔
classification
Cost-sensitive
✔ ✔ ✔
learning method
Deep Transfer
Learning-based ✔
System
MicFoal ✔
Deep Reinforce-
ment Learning- ✔
based System
Borderline-
✔
DEMNET
Adaptive Weight-
✔ ✔
ing Mechanism
THPIA Algorithm ✔ ✔
Belief Function &
✔ ✔ ✔
Ensemble
SDCU ✔ ✔
GLOS ✔ ✔
HCE-MCD ✔ ✔
Random Balance ✔ ✔ ✔ ✔ ✔ ✔

Table 1.  The state-of-the-art methods for imbalanced multiclass classification with their respective properties
marked with a checkmark (✓).

Initially, dataset D is defined as a collection of sample-label pairs, represented as {(X, Y )|XǫRN×M , Y ǫRN×M },
where X represents the set of input features, and Y represents the corresponding labels. To begin the process,

Scientific Reports | (2025) 15:3460 | https://doi.org/10.1038/s41598-024-84786-2 5

Vol.:(0123456789)
 www.nature.com/scientificreports/

Fig. 1.  The process of the proposed HCBOU algorithm.

Notation Size Description

N 1 The number of instances in the training set
Nci 1 The number of instances in class i
M 1 The number of attributes
C 1 The number of classes
X N ×M The set of features
Y N ×C The set of labels
α 1 The number of majority classes (the number of datasets in Dmaj)
β 1 The number of minority classes (the number of datasets in Dmin)
Di Nci × (M + 1) The subset of training set containing the class i
Dmaj Nci × (M + 1) × α Majority set
Dmin Nci × (M + 1) × β Minority set
S 1 The number of instances of each class after balancing
i Nci × (M + 1) The dataset containing cluster centres for dataset i
O 1 Optimal number of clusters for minority class clustering
i,j
C 1 The number of instances in cluster i of dataset j
nc′ 1 The number of clusters with sufficient number of instances
C′ 1 × nc′ The list of clusters with sufficient number of instances
m 1 Minimum number of instances in each cluster
ei,j̇ 1 The total absolute error (TAE) of the distance between instance i and cluster j
lei 1 × nc′ The list of TAE of the distance between instance i and other clusters
mei 1 The cluster with the smallest TAE value for instance i
Wi,j 1 The weight of cluster i in the dataset j

Si,j 1 The number of instances generated in cluster i of dataset j
K 1 The number of nearest neighbours in SMOTE
D′ (S × C) × (M + 1) Balanced dataset
σi 1×M The centre of cluster i
Di,j,k 1×M The sample i from the cluster j of dataset k

Table 2.  The notations used in the proposed algorithm.

Scientific Reports | (2025) 15:3460 | https://doi.org/10.1038/s41598-024-84786-2 6

Vol:.(1234567890)
www.nature.com/scientificreports/

the first step is to calculate the desired number of instances for each class, which is determined using Eq. (1), as
shown below. Based on this, the dataset is divided into majority and minority groups using Eqs. (2)-(3):
N
S= (1)
C
 
D|class = i, (Nci ≥ S) ∈ Dmaj (2)
 
D|class = i, (Nci < S) ∈ Dmin (3)
Based on these equations, if the subset containing class i has more instances than S , it is placed in the major-
ity class ( Dmaj ); otherwise, it is placed in the minority class ( Dmin). This clear classification allows us to focus on
addressing the imbalance more effectively.
To handle the issue of imbalanced data and noise, clustering is applied in conjunction with undersampling.
In this context, clustering groups data points based on their similarity, with the goal of maximizing intra-cluster
similarity and minimizing inter-cluster similarity. Such clustering techniques help in forming coherent subgroups
within the data, allowing for more targeted data manipulation. Instead of relying solely on undersampling, we
employ a clustering strategy that ensures representative data points are retained. The centers of these clusters are
used to represent the data, thereby reducing the risk of losing important information due to undersampling. This
step is crucial because it helps to maintain data integrity while reducing the impact of noise and redundancy.
Additionally, the use of clustering-based undersampling also allows for better handling of high-dimensional
data, where traditional methods might struggle to capture the inherent structure of the dataset. However, it is
important to note that sometimes this approach may not be effective if the clusters are not well-separated or
if significant overlap exists between them. In those cases, adjustments are made to ensure proper separation
between clusters. Figure 2 illustrates this approach, where data points are reduced from 50 to 10 while preserv-
ing the essential patterns.
Clustering is a fundamental step in our approach, as it allows us to effectively group data points, minimize
intra-cluster variance, and optimize the sampling process, which is more critical than the specific clustering
technique itself. Among various clustering methods, K-means is chosen because of its simplicity, efficiency, and
suitability for large-scale datasets. The K -means clustering algorithm produces S groups for each subset of (Dmaj ).
The centers of each cluster are then used as undersampled data. This clustering-based undersampling approach
ensures that we retain representative data, even from large majority classes, without losing valuable informa-
tion. For each subset of (Dmin ), high-quality synthetic data are generated using a combination of K -means and
oversampling. Here, we apply SMOTE in a localized manner within each cluster, which significantly reduces
computation while improving the quality of synthetic data. Localized oversampling is particularly beneficial in
cases where global oversampling would introduce noise or obscure meaningful patterns. This method ensures
that the minority class is better represented, and the underlying data relationships are preserved. By focusing on
localized cluster regions, this approach helps to mitigate the risk of generating redundant or irrelevant samples,
which is a common issue in traditional oversampling techniques. By generating synthetic data for each cluster
individually, we also avoid overfitting, a common risk when oversampling is applied to the entire dataset. In this
regard, the well-known silhouette a­ lgorithm43 is used to find the best number of clusters (O) for the K -means
algorithm. Given that some clusters may contain a small number of instances after clustering, this could lead

Fig. 2.  Illustration of using the cluster centres as replacements for the original data.

Scientific Reports | (2025) 15:3460 | https://doi.org/10.1038/s41598-024-84786-2 7

Vol.:(0123456789)
 www.nature.com/scientificreports/

to the generation of low-quality data. To maintain consistency and avoid small, non-representative clusters, we
establish a minimum threshold for the number of instances per cluster. To address this problem, at first, the
clusters that have at least m instances are placed in C′ . The clusters with data points fewer than the threshold m
are considered sparse clusters. Then, the data points from sparse clusters are redistributed to the nearest cluster
from the list of C′ . The redistribution of sparse cluster data not only enhances the quality of generated samples
but also helps in maintaining a smooth decision boundary. An illustration of data redistribution from sparse
clusters is shown in Fig. 3.
Next, we calculate the appropriate amount of data to be generated for each cluster in (Dmin). The weight of the
jth cluster belonging to the i th subgroup of (Dmin ) is denoted by Wj,Dmin [i]. Once the weight is determined, it can be
used to ascertain the adequate number of samples for generation in the jth cluster belonging to the i th subgroup
of Dmin , denoted by  Sj,Dmin [i]. The calculation of Wj,Dmin [i] and 
Sj,Dmin [i] is shown in Eqs. (4) and (5), respectively.
m,D [j]
C
Wm,Dmin [j] = m min ; m ∈ {1,2, . . . , O} (4)
m,Dmin [j]
C 1

 m,D j )⌋; m ∈ {1,2, . . . , O}

Sm,Dmin [j] = ⌊W m,Dmin [j] × (S − C (5)
min [ ]

To address the disparity between the available data for each class and the desired number of samples per
class, it is necessary to generate additional data. The weight of each cluster reflects the relative importance of the
related cluster in the process of data augmentation. After determining the required size of data for generation,
the SMOTE method can be used to generate the data. The dynamic nature of cluster-weighted data generation
ensures that the synthetic data is well-distributed, reducing bias and enhancing the diversity of samples. One
of the required parameters of the problem is K , which represents the number of neighbors considered for data
generation. Finally, the balanced dataset is reconstructed, and predictive models are implemented using ensemble
approaches. These ensemble models leverage OVO and OVA decomposition schemes, further enhancing the
predictive power of the model by combining results from multiple binary classifiers. Algorithm 1 represents the
pseudo-code of the proposed algorithm.

Fig. 3.  Illustration of data redistribution from sparse clusters.

Scientific Reports | (2025) 15:3460 | https://doi.org/10.1038/s41598-024-84786-2 8

Vol:.(1234567890)
www.nature.com/scientificreports/

Input: The training set = ( , ); = { 1 , 2 , … , }; = { 1 , 2 , … , }, ⴜ,

Output: ′ obtained after balancing
←∅ // Create new list ( )
←∅ // Create new list ( )
for ← 1, . . . , do
if ≥ then
← // Append dataset ( ) to list ( )
else
← // Append dataset ( ) to list ( )

← // Assign value to
for ← 1, . . . , α do
←∅
function -Means ( [ ], = ): // [ ] refers to th dataset in
[ ] ← -Means cluster centers
[]←∅ // Remove all instances from []
[]← [] // Append [ ] to []
for ← 1, . . . , do
← function silhouette ( [ ]) // Using silhouette method to find optimal cluster number
function -Means ( [ ], = )
′
←∅ // Create new list ( ′ )
for ← 1, . . . , do // Moving instances from sparse clusters to their nearest cluster
if ̃ , [ ] ≥ ⴜ then
′
←
for ← 1, . . . , do
if ̃ , [ ] < ⴜ then
← −1
for ← 1, … , ̃ , [ ] do
←∅
for ← ( ′ ) do

, =∑ | , , [ ] − |
=1
← ,
← ( )
th
append the instances to the cluster
for ← 1, . . . , do
̃ , []
, [] = ∑ =1 ̃ ,
// Determining weight of each cluster
[]
̂, [] =⌊ , []×( − ,
̃ [ ] )⌋ // Determining the number of samples generated
function SMOTE ( [ ], = ̂ , [ ], ℎ = )

′
← function Concatenate ( , )
1 ← Bagging-classifier ( ′ )
2 ← Boosting-classifier ( ′ )

Algorithm 1:.  Hybrid cluster-based oversampling and undersampling for multiclass classification

Experiments
This section evaluates the performance and validates the efficiency of our proposed approach in addressing
multiclass imbalanced learning challenges. We will begin this section by providing a concise overview of the
datasets, evaluation metrics, and comparison algorithms. Afterward, we will present the analysis and results of
the experiments. We utilized the Python imbalanced-learn l­ ibrary44 to deal with imbalanced data.

Datasets
The experiments involve 30 datasets with imbalanced classes, which were obtained from online repositories
such as UCI, (https://​archi​ve.​ics.​uci.​edu/​datas​ets) KEEL (https://​sci2s.​ugr.​es/​keel/​datas​ets.​php) and OpenML.
(https://​www.​openml.​org/​search?​type=​data&​sort=​runs&​status=​active) The imbalance ratio (IR), which is used
to illustrate the degree of data imbalance, is defined as follows:
max(Nci )
IR = ; i = 1, 2, . . . , C (6)
min(Nci )
Table 3 provides an overview of these datasets. Each of the datasets contains at least three classes, and the
imbalance ratio of the multiclass datasets ranges from 1.09 to 2160.

Scientific Reports | (2025) 15:3460 | https://doi.org/10.1038/s41598-024-84786-2 9

Vol.:(0123456789)
 www.nature.com/scientificreports/

Datasets #Ex IR #A #Nu #No #C Class distribution

Arrhythmia 452 122.5 279 206 63 13 245, 50, 44, 25, 22, 15, 15, 13, 9, 5, 4, 3, 2
Autos 159 16.00 25 15 10 6 48, 46, 29, 20, 13, 3
Balance 625 5.87 4 4 0 3 288, 288, 49
Cardiotocography-10classes 2126 10.92 35 35 0 10 579, 384, 332, 252, 197, 107, 81, 72, 69, 53
Cardiotocography-3classes 2126 9.40 35 35 0 3 1655, 295, 176
Contraceptive 1473 1.88 9 6 3 3 629, 511, 333
Dermatology 366 5.60 34 34 0 6 112, 72, 61, 52, 49, 20
Ecoli 336 71.50 7 7 0 8 143, 77, 52, 35, 20, 5, 2, 2
Flags 194 15.00 28 2 26 8 60, 40, 36, 27, 15, 8, 4, 4
Glass 214 8.44 9 9 0 6 76, 70, 29, 17, 13, 9
Heart-cleveland 303 12.61 13 13 0 5 164, 55, 36, 35, 13
Heart-switzerland 123 9.60 12 12 0 5 48, 32, 30, 8, 5
Led7digit 500 1.54 7 0 7 10 57 57 53 52 52 51 49 47 45 37
Lenses 24 3.75 4 0 4 3 15, 5, 4
Lymphography 148 40.5 18 3 15 4 81, 61, 4, 2
Molec-biol-splice 3190 2.15 60 0 60 3 1655, 768, 767
New-thyroid 215 5.00 5 5 0 3 150, 35, 30
Nursery 12,960 2160 8 0 8 5 4320, 4266, 4044, 328, 2
Pageblocks 548 164.0 10 10 0 5 492, 33, 12, 8, 3
Penbased 1100 1.09 16 16 0 10 115, 115, 114, 114, 114, 106, 106, 106, 105, 105
Shuttle 2175 853.0 9 9 0 5 1706, 338, 123, 6, 2
Statlog-landsat 6435 2.44 36 36 0 6 1533, 1508, 1358, 707, 703, 626
Steel-plates 1941 12.23 27 27 0 7 673, 402, 391, 190, 158, 72, 55
Thyroid 720 39.17 21 6 15 3 666, 37, 17
Vehicle 846 1.09 18 18 0 4 218, 217, 212, 199
Vertebral-column-3classes 310 2.50 6 6 0 3 150, 100, 60
Wine 178 1.47 13 13 0 3 71, 59, 48
Winequality-red 1599 68.10 11 11 0 6 681, 638, 199, 53, 18, 10
Yeast 1484 92.60 8 8 0 10 463, 429, 244, 163, 51, 44, 35, 30, 20, 5
Zoo 101 10.25 16 0 16 7 41, 20, 13, 10, 8, 5, 4

Table 3.  30 benchmark datasets are described along with their properties (#Ex: examples, #A: attributes, #Nu:
numerical feature, #No: nominal feature, #C: classes).

Evaluation metrics
Accuracy is a widely used metric to assess the effectiveness of a classification model. In the case of multiclass
classification, accuracy can be defined as follows:
 C
1
Accuracy = × TP i (7)
N
i=1

where TP i stands for true positive of the i th class, which means a positive sample is correctly identified as posi-
tive by the model. Also, N is the total number of instances in the training set, and C is the number of classes.
It should be noted that, in multiclass classification problems, the accuracy metric can be a misleading perfor-
mance metric. Accuracy does not consider class distribution and favors the majority class, whereas a classifier
that always predicts the majority class would still achieve a high accuracy. The F1-score45 is commonly used to
evaluate multiclass classification models, where this metric is calculated per class and then averaged for an overall
performance measure. This metric is a harmonic mean of precision and recall, and it is calculated for each class
separately based on the following equation:
 
1 + β 2 ×Precisioni × Recall i
F1 − scorei = (8)
β 2 ×Precisioni + Recall i
Then, the arithmetic mean of all F1 − scorei becomes the model’s final F1-score, defined in Eq. (9):
C
1 
F1 = × F1 − scorei (9)
C
i=1

where β is utilized to balance the significance of precision and recall. In addition, recall and precision are defined
as follows:

Scientific Reports | (2025) 15:3460 | https://doi.org/10.1038/s41598-024-84786-2 10

Vol:.(1234567890)
www.nature.com/scientificreports/

TPi
Precisioni = (10)
TPi + FPi

TPi
Recalli = (11)
TPi + FNi
where FP i stands for false positive of the i th class, which means a negative sample wrongly identified as
positive by the model, and FN i stands for false negative of the i th class, which means a positive sample wrongly
identified as negative by the model.
Averaged-precision and averaged-recall46 are measures used in multiclass classification problems to calculate
the average precision and recall values across all classes.
C
1 
Averaged − precision = × Precisioni (12)
C
i=1

C
1 
Averaged − recall = × Recall i (13)
C
i=1

The G-mean ­score47 is another metric that can be used to assess the overall effectiveness of a multiclass clas-
sifier in a more balanced way. The G-mean in multiclass classification is calculated as the geometric mean of the
recall scores for all the classes based on the following equation:
 C
 C1

G − mean = Recall i (14)
i=1

The MAUC (Mean Area Under the Curve)48 is a measure that calculates the average AUC for all possible pairs
of classes in a multiclass classification problem. This measure is defined in Eq. (15):
 C
1
MAUC = × AUC(i, j)
C(C − 1) (15)
i,j=1
i�=j

where AUC(i, j) represents the area under the curve that corresponds to the pair of classes i and j.

Parameter setting
In this study, we use a grid search technique to find the hyperparameter values that maximize the model’s
performance. Considering K and m as two hyperparameters of the proposed model, the search space for these
hyperparameters is {3, 4, 5} and {3, 4, . . . , 15}, respectively. The performance of the proposed model is evaluated
for all combinations of K and values in the search space, and the hyperparameters that yield the best results are
selected for the final model.
The performance of the proposed HCBOU is evaluated and compared to various multiclass imbalance learn-
ing techniques, including EasyEnsemble, RUSBoost, SMOTEBagging, Roughly Balanced Bagging, AdaBoost.NC
and SMOTE + AdaBoost, all of which have demonstrated effective performance in handling imbalanced data.
The parameter configurations of these approaches are represented in Table 4.

Comparing imbalance learning approaches

The proposed method is compared with the methods EasyEnsemble, RUSBoost, Balanced Random Forest,
Roughly Balanced Bagging, AdaBoost.NC and SMOTE + AdaBoost. The comparison is performed using both
OVO and OVA classification schemes. Table 5 shows the proposed model and competing methods, along with
their classification schemes. The proposed HCBOU is implemented using bagging and boosting approaches.
The G-mean, F1-score, averaged precision, averaged recall, and MAUC metrics are used for comparisons across
30 datasets.

Results
The average rank of the methods in terms of six predefined metrics based on 30 datasets is presented in Table 6.
The 95% confidence interval of the average ranks is also shown in this table ( α = 0.05). The methods have been
sorted from best to worst based on the average rank. Based on the results, it can be observed that the proposed
method, in combination with bagging and OVA classification scheme (OVA-HCBOUBag), provides the best
performance in terms of all evaluation metrics, except for the accuracy metric, where it ranks second. In addition,
the proposed method in combination with bagging demonstrates better results compared to its combination with
boosting, and the proposed method under the OVA classification scheme shows better performance compared
to the OVO classification scheme.
To evaluate the significance of the differences between the OVA-HCBOUBag (the proposed method with
the best performance) and other methods, a Wilcoxon signed-rank ­test49 at the significance level of α = 0.05 is
performed on six performance metrics where the results are presented in Table 7. The Wilcoxon signed-rank

Scientific Reports | (2025) 15:3460 | https://doi.org/10.1038/s41598-024-84786-2 11

Vol.:(0123456789)
 www.nature.com/scientificreports/

Method Parameters
estimator = AdaBoost;
EasyEnsemble no. of estimators = 10;
replacement = false
no. of estimators = 50;
replacement = false;
RUSBoost learning rate = 1;
estimator = Decision tree;
estimator_depth ∈ {1,2, 3}
no. of estimators = 100;
replacement = false;
Balanced Random Forest estimator = Decision tree
estimator_depth ∈ {1,2, 3}

max _feature = no.offeature
no. of estimators = 10;
estimator = Decision tree;
Roughly Balanced Bagging
estimator_depth ∈ {1,2, 3}
bootstrap = true
no. of estimators = 100;
learning rate = 1;
AdaBoost.NC
estimator = Decision tree;
estimator_depth {1,2, 3}
estimator = AdaBoost;
no. of estimators = 50;
SMOTE + AdaBoost
learning rate = 1;
SMOTE_ k_neighbors ∈{3,4, 5}
SMOTE_ k_neighbors ∈ {3,4, 5}
m {3,4, . . . , 15}
Bagging:
no. of estimators = 100;
estimator = Decision tree;
HCBOU estimator_depth ∈ {1,2, 3}
Boosting:
no. of estimators = 50;
estimator = Decision tree;
estimator_depth ∈ {1,2, 3}
learning rate = 1

Table 4.  Parameter settings for ensemble methods. Grid search is used to determine the optimal values of
their parameters.

Classification scheme
Method OVA OVO
EasyEnsemble OVA-EE OVO-EE
RUSBoost OVA-RUS OVO-RUS
Balanced Random Forest OVA-BRF OVO-BRF
Roughly Balanced Bagging OVA-RBB OVO-RBB
AdaBoost.NC OVA-Ada.NC OVO-Ada.NC
SMOTE + AdaBoost OVA-S + Ada OVO-S + Ada
HCBOU-Bagging OVA-HCBOUBag OVO-HCBOUBag
HCBOU-Boosting OVA-HCBOUBoo OVO-HCBOUBoo

Table 5.  Ensemble methods used along with their classification schemes.

test is a non-parametric statistical test for comparing two methods and is appropriate when data violates the
assumptions of normality or equal variances. The null hypothesis for the Wilcoxon signed-rank test states that
there is no significant difference between the two related groups being compared. The statistical analysis showed
that the proposed method significantly improved all metrics compared to the other methods (p-value < 0.05).
For a more comprehensive comparison between the proposed method and other competing methods, Table 8
shows the pair-wise performance of the methods against each other. In this table, each cell shows the number of
superiorities of the methods shown in rows over those shown in columns. According to the results, the proposed
method resulted in superior effectiveness in comparison to the competing methods in terms of all performance
measures and in most of the datasets. Furthermore, the proposed algorithm has demonstrated better performance
using the bagging approach under most scenarios.
Table 9 presents a comparison of the performance of OVO and OVA schemes across various ensemble
methods and evaluation metrics. The results indicate that the performance of the OVO is generally better than
the OVA. This is justifiable due to the higher number of models in the OVO scheme. However, in some cases, it
can lead to overfitting and inferior performance compared to the OVA scheme. Based on the evaluation of 30

Scientific Reports | (2025) 15:3460 | https://doi.org/10.1038/s41598-024-84786-2 12

Vol:.(1234567890)
www.nature.com/scientificreports/

Accuracy Averaged-precision Averaged-recall F1 G-Mean MAUC​

Algorithm Rank Algorithm Rank Algorithm Rank Algorithm Rank Algorithm Rank Algorithm Rank
OVA-HCBOU- OVA-HCBOU- OVA-HCBOU- OVA-HCBOU- OVA-HCBOU-
OVA-BRF 5.80 ± 1.5 5.90 ± 1.5 5.30 ± 1.5 5.40 ± 1.6 5.03 ± 1.6 5.10 ± 1.5
Bag Bag Bag Bag Bag
OVA-HCBOU- OVO- OVO- OVO- OVO- OVO-
6.30 ± 1.5 6.13 ± 2.0 5.37 ± 1.9 5.53 ± 2.0 5.47 ± 1.9 5.37 ± 1.9
Bag HCBOUBag HCBOUBag HCBOUBag HCBOUBag HCBOUBag
OVA-Ada.NC 6.47 ± 1.6 OVO-Ada.NC 7.40 ± 1.4 OVA-BRF 7.17 ± 1.5 OVA-BRF 7.00 ± 1.5 OVO-EE 6.87 ± 1.3 OVA-BRF 6.83 ± 1.5
OVO-S + Ada 6.63 ± 1.7 OVA-BRF 7.43 ± 1.5 OVO-EE 7.57 ± 1.6 OVO-S + Ada 7.53 ± 1.5 OVA-EE 7.53 ± 1.5 OVO-EE 7.63 ± 1.6
OVA-HCBOU- OVA-HCBOU-
OVO-Ada.NC 7.03 ± 1.3 OVO-S + Ada 7.63 ± 1.6 OVA-S + Ada 7.87 ± 1.5 OVA-S + Ada 7.70 ± 1.6 7.53 ± 1.7 8.03 ± 1.5
Boo Boo
OVA-HCBOU-
OVA-S + Ada 7.07 ± 1.6 OVA-Ada.NC 7.90 ± 1.7 7.93 ± 1.5 OVO-Ada.NC 7.90 ± 1.4 OVA-BRF 7.80 ± 1.5 OVO-S + Ada 8.20 ± 1.8
Boo
OVO- OVA-HCBOU- OVA-HCBOU- OVO-
7.07 ± 1.8 7.93 ± 1.7 OVO-S + Ada 8.10 ± 1.7 8.17 ± 1.7 7.90 ± 1.7 OVA-S + Ada 8.40 ± 1.7
HCBOUBag Boo Boo HCBOUBoo
OVO- OVO-
OVO-EE 9.07 ± 1.5 OVA-S + Ada 8.00 ± 1.8 8.27 ± 1.8 OVA-Ada.NC 8.50 ± 1.4 OVO-BRF 8.20 ± 1.4 8.53 ± 1.8
HCBOUBoo HCBOUBoo
OVO-
OVA-RBB 9.20 ± 1.5 OVO-BRF 8.57 ± 1.6 OVO-BRF 8.77 ± 1.7 8.63 ± 1.7 OVA-S + Ada 8.97 ± 1.7 OVO-BRF 8.70 ± 1.6
HCBOUBoo
OVO-
OVO-BRF 9.40 ± 1.7 8.67 ± 1.8 OVO-Ada.NC 9.37 ± 1.4 OVO-BRF 8.80 ± 1.7 OVO-S + Ada 9.10 ± 2.0 OVO-Ada.NC 9.23 ± 1.5
HCBOUBoo
OVO-
9.90 ± 1.5 OVO-EE 9.23 ± 1.7 OVA-Ada.NC 9.50 ± 1.4 OVO-EE 9.07 ± 1.6 OVO-RBB 9.53 ± 1.7 OVA-Ada.NC 9.33 ± 1.5
HCBOUBoo
OVA-HCBOU-
10.00 ± 1.6 OVA-RBB 9.27 ± 1.3 OVA-EE 9.53 ± 1.9 OVA-RBB 9.77 ± 1.4 OVA-RUS 9.80 ± 1.7 OVA-EE 9.40 ± 1.8
Boo
OVO-RBB 10.37 ± 1.5 OVA-RUS 10.07 ± 1.8 OVO-RBB 10.03 ± 1.8 OVA-RUS 10.20 ± 1.9 OVA-RBB 9.93 ± 1.2 OVO-RUS 10.17 ± 1.6
OVA-RUS 10.43 ± 1.9 OVA-EE 10.37 ± 2.0 OVA-RUS 10.20 ± 2.0 OVO-RBB 10.23 ± 1.5 OVO-Ada.NC 10.60 ± 1.7 OVA-RUS 10.20 ± 1.9
OVA-EE 10.60 ± 2.0 OVO-RUS 10.70 ± 1.5 OVA-RBB 10.50 ± 1.4 OVO-RUS 10.70 ± 1.5 OVO-RUS 10.80 ± 1.6 OVO-RBB 10.40 ± 1.7
OVO-RUS 10.67 ± 1.5 OVO-RBB 10.80 ± 1.4 OVO-RUS 10.53 ± 1.6 OVA-EE 10.87 ± 2.0 OVA-Ada.NC 10.93 ± 1.6 OVA-RBB 10.47 ± 1.3

Table 6.  Average rank of performance metrics for the methods on 30 datasets. The proposed method in italic.

Algorithm Accuracy Averaged-Precision Averaged-recall F1 G-Mean MAUC​

OVA-EE 9.45E-03 ( >) 7.43E-03 ( >) 1.23E-02 ( >) 2.31E-03 ( >) 6.96E-02 ( =) 8.38E-03 ( >)
OVO-EE 4.68E-02 ( >) 2.11E-02 ( >) 9.87E-02 ( =) 1.01E-02 ( >) 8.00E-02 ( =) 5.27E-02 ( =)
OVA-RBB 2.10E-02 ( >) 3.55E-03 ( >) 9.52E-05 ( >) 2.64E-04 ( >) 8.74E-05 ( >) 6.77E-05 ( >)
OVO-RBB 3.02E-03 ( >) 5.11E-04 ( >) 7.20E-04 ( >) 7.05E-04 ( >) 1.46E-03 ( >) 2.37E-04 ( >)
OVA-BRF 5.77E-01 ( =) 1.84E-01 ( =) 5.53E-02 ( =) 1.55E-01 ( =) 1.59E-02 ( >) 6.51E-02 ( =)
OVO-BRF 1.46E-02 ( >) 4.80E-02 ( >) 2.23E-02 ( >) 2.17E-02 ( >) 1.56E-02 ( >) 1.20E-02 ( >)
OVA-RUS 6.18E-03 ( >) 6.17E-03 ( >) 3.03E-03 ( >) 1.94E-03 ( >) 1.62E-03 ( >) 1.81E-03 ( >)
OVO-RUS 2.90E-03 ( >) 1.86E-03 ( >) 7.45E-04 ( >) 8.78E-04 ( >) 4.77E-04 ( >) 8.45E-04 ( >)
OVA-S + Ada 6.21E-01 ( =) 1.38E-01 ( =) 3.06E-02 ( >) 6.51E-02 ( =) 6.16E-03 ( >) 9.40E-03 ( >)
OVO-S + Ada 8.45E-01 ( =) 1.52E-01 ( =) 3.06E-02 ( >) 7.00E-02 ( =) 6.76E-03 ( >) 1.74E-02 ( >)
OVA-Ada.NC 8.53E-01 ( =) 4.79E-02 ( >) 5.77E-04 ( >) 5.97E-03 ( >) 1.51E-04 ( >) 5.85E-04 ( >)
OVO-Ada.NC 5.71E-01 ( =) 2.02E-01 ( =) 2.15E-03 ( >) 3.57E-02 ( >) 9.15E-04 ( >) 1.47E-03 ( >)

Table 7.  Results of Wilcoxon signed-rank tests comparing the proposed OVA-HCBOUBag method with other
methods applied on 30 datasets, where the ( =) sign denotes no significant difference between the compared
techniques, while ( >) indicates that the proposed method outperforms the compared method.

datasets and different performance measures, it has been observed that HCBOU-Bagging demonstrates superior
performance when employed in conjunction with the OVO scheme, whereas HCBOU-Boosting yields better
results when utilizing the OVA scheme.
Finally, Fig. 4 illustrates the relationship between the imbalance ratio (IR) and performance metrics using
scatter plots and regression analysis. This plot can offer valuable insights into how the imbalance ratio affects
performance metrics and can help identify any trends or patterns within the data. As the degree of class imbalance
increases, the performance metrics remain relatively stable, suggesting that the method is consistently effective
across various ratios of imbalance. However, some dataset performances may deviate from the fitted line due to
internal variations within the data.

Scientific Reports | (2025) 15:3460 | https://doi.org/10.1038/s41598-024-84786-2 13

Vol.:(0123456789)
 (a)
OVA- OVO- OVO- OVA- OVO- OVA- OVO- OVA- OVO- OVO- OVA- OVA- OVA- OVO-

Vol:.(1234567890)
HCBOUBag HCBOUBag HCBOUBoo HCBOUBoo RBB RBB S+Ada Ada.NC RUS Ada.NC BRF OVO-EE RUS S+Ada BRF OVA-EE
OVA-HCBOUBag 20 26 26 20 20 13 15 24 14 14 22 20 15 21 21
OVO-HCBOUBag 20 23 23 20 18 15 12 24 14 13 20 20 16 20 20

Scientific Reports |
OVO-HCBOUBoo 8 11 18 17 14 8 7 18 8 9 16 16 11 15 19
OVA-HCBOUBoo 9 10 16 14 13 9 10 18 9 8 17 17 9 12 17
OVO-RBB 10 10 15 16 12 8 6 16 7 7 15 14 8 14 18
OVA-RBB 10 13 17 18 21 10 7 20 8 8 17 17 10 16 18
OVO-S+Ada 17 16 23 21 23 21 16 24 18 16 21 24 22 21 25
OVA-Ada.NC 16 19 23 22 25 24 18 26 20 19 25 24 20 21 26

(2025) 15:3460 |
OVO-RUS 7 7 12 13 15 11 11 8 8 8 19 23 10 16 22
OVO-Ada.NC 17 17 23 23 24 23 19 18 27 18 25 25 17 23 27
OVA-BRF 18 18 22 23 24 24 17 18 25 18 24 23 19 26 26
www.nature.com/scientificreports/

OVO-EE 9 11 14 15 16 14 10 11 20 11 11 22 9 19 25
OVA-RUS 11 11 14 14 17 14 9 10 12 10 13 13 10 17 19
OVA-S+Ada 15 15 21 21 23 22 16 12 22 17 16 23 25 21 24
OVO-BRF 10 11 15 19 17 15 10 13 18 11 12 17 20 12 24
OVA-EE 10 11 11 14 13 14 7 9 12 8 10 10 15 8 13
(b)
OVA- OVO- OVO- OVA- OVO- OVA- OVO- OVA- OVO- OVO- OVA- OVA- OVA- OVO-
HCBOUBag HCBOUBag HCBOUBoo HCBOUBoo RBB RBB S+Ada Ada.NC RUS Ada.NC BRF OVO-EE RUS S+Ada BRF OVA-EE
OVA-HCBOUBag 15 25 24 24 26 22 23 23 23 19 19 20 18 19 21
OVO-HCBOUBag 20 26 22 22 24 22 23 23 22 22 19 20 20 19 20
OVO-HCBOUBoo 6 6 12 18 18 17 21 20 18 15 16 17 17 17 18
OVA-HCBOUBoo 9 10 19 21 21 16 18 20 17 13 16 19 14 16 18
OVO-RBB 6 8 12 9 18 12 13 16 14 12 13 14 9 13 14
OVA-RBB 4 6 12 9 14 11 13 17 13 11 11 14 10 10 12
OVO-S+Ada 8 9 14 14 19 20 19 22 19 16 17 22 19 18 21

https://doi.org/10.1038/s41598-024-84786-2
OVA-Ada.NC 8 8 9 13 18 18 14 20 19 13 17 21 15 16 20
OVO-RUS 8 8 10 11 15 14 11 14 10 9 11 21 11 14 19
OVO-Ada.NC 8 9 12 14 17 18 15 18 24 15 17 20 15 18 21
OVA-BRF 12 9 15 18 19 20 17 22 24 19 19 21 19 21 22
OVO-EE 12 12 14 15 18 20 14 18 23 18 14 22 15 20 26
OVA-RUS 11 11 13 12 17 17 9 12 13 13 13 11 13 14 15
OVA-S+Ada 12 11 14 16 22 21 16 17 20 19 15 17 19 21 22
OVO-BRF 12 12 13 15 18 21 13 18 19 15 15 14 20 12 21
OVA-EE 10 11 12 13 17 19 11 13 14 12 12 9 19 10 15
(c)
OVA- OVO- OVO- OVA- OVO- OVA- OVO- OVA- OVO- OVO- OVA- OVA- OVA- OVO-
HCBOUBag HCBOUBag HCBOUBoo HCBOUBoo RBB RBB S+Ada Ada.NC RUS Ada.NC BRF OVO-EE RUS S+Ada BRF OVA-EE
OVA-HCBOUBag 16 24 21 26 23 18 19 23 18 19 20 20 18 18 20
OVO-HCBOUBag 18 20 21 23 21 20 18 23 18 20 20 20 18 20 20
Continued

14
 (c)
OVA- OVO- OVO- OVA- OVO- OVA- OVO- OVA- OVO- OVO- OVA- OVA- OVA- OVO-
HCBOUBag HCBOUBag HCBOUBoo HCBOUBoo RBB RBB S+Ada Ada.NC RUS Ada.NC BRF OVO-EE RUS S+Ada BRF OVA-EE
OVO-HCBOUBoo 7 12 13 22 17 14 14 18 12 14 16 19 14 15 17
OVA-HCBOUBoo 12 11 18 22 18 15 17 20 13 15 17 20 14 16 19

Scientific Reports |
OVO-RBB 4 7 8 8 13 11 9 15 7 10 16 14 9 13 16
OVA-RBB 7 9 13 12 19 13 13 18 8 13 17 14 13 16 19
OVO-S+Ada 12 11 17 15 20 18 17 23 17 17 19 22 19 19 23
OVA-Ada.NC 12 13 16 14 22 18 16 24 19 18 21 21 18 19 24
OVO-RUS 8 8 12 11 16 13 10 9 8 9 14 20 11 12 21
OVO-Ada.NC 13 13 18 18 24 23 17 17 26 20 24 24 17 19 25

(2025) 15:3460 |
OVA-BRF 12 11 16 16 21 18 15 17 24 14 23 20 17 21 24
OVO-EE 11 11 14 14 15 14 12 13 19 11 10 19 13 13 22
OVA-RUS 11 11 11 11 17 17 9 12 15 9 14 14 12 16 20
www.nature.com/scientificreports/

OVA-S+Ada 12 13 17 16 22 18 16 14 20 16 17 19 20 16 22
OVO-BRF 13 11 15 15 18 15 12 15 21 14 16 21 18 17 23
OVA-EE 11 11 13 12 15 12 8 9 13 8 10 13 15 10 12
(d)
OVA- OVO- OVO- OVA- OVO- OVA- OVO- OVA- OVO- OVO- OVA- OVA- OVA- OVO-
HCBOUBag HCBOUBag HCBOUBoo HCBOUBoo RBB RBB S+Ada Ada.NC RUS Ada.NC BRF OVO-EE RUS S+Ada BRF OVA-EE
OVA-HCBOUBag 16 24 25 25 25 21 22 25 20 19 21 19 17 18 21
OVO-HCBOUBag 18 23 22 24 22 22 23 23 21 21 21 19 19 19 21
OVO-HCBOUBoo 7 9 13 18 19 13 19 20 13 14 15 17 15 16 17
OVA-HCBOUBoo 8 10 18 20 18 14 18 19 14 14 19 20 15 15 18
OVO-RBB 5 6 12 10 16 12 10 16 9 10 15 14 9 14 19
OVA-RBB 5 8 11 12 16 12 12 17 10 12 16 16 12 13 17
OVO-S+Ada 9 9 18 16 19 19 18 23 18 16 20 22 20 21 24
OVA-Ada.NC 9 8 11 13 21 19 15 25 18 16 20 20 18 19 24
OVO-RUS 6 8 10 12 15 14 10 8 8 9 13 21 10 13 23

https://doi.org/10.1038/s41598-024-84786-2
OVO-Ada.NC 11 10 17 17 22 21 16 18 26 17 24 23 16 20 25
OVA-BRF 12 10 16 17 21 19 16 19 24 17 23 22 19 22 25
OVO-EE 10 10 15 12 16 15 11 13 20 11 10 20 11 17 25
OVA-RUS 12 12 13 11 17 15 9 13 13 10 12 13 10 15 19
OVA-S+Ada 13 12 16 15 22 19 15 14 21 17 15 21 22 21 24
OVO-BRF 13 12 14 16 17 18 10 15 20 13 14 17 19 12 26
OVA-EE 10 10 13 13 12 14 7 9 10 8 9 10 15 8 9
(e)
OVA- OVO- OVO- OVA- OVO- OVA- OVO- OVA- OVO- OVO- OVA- OVA- OVA- OVO-
HCBOUBag HCBOUBag HCBOUBoo HCBOUBoo RBB RBB S+Ada Ada.NC RUS Ada.NC BRF OVO-EE RUS S+Ada BRF OVA-EE
OVA-HCBOUBag 17 25 23 24 25 22 24 23 23 21 21 21 18 20 22
OVO-HCBOUBag 20 25 21 22 23 22 24 23 22 22 20 20 20 19 20
OVO-HCBOUBoo 9 9 13 20 21 18 20 20 18 17 17 18 18 16 19
OVA-HCBOUBoo 11 12 19 22 22 16 18 19 17 15 17 20 17 16 20
Continued

15

Vol.:(0123456789)
 (e)
OVA- OVO- OVO- OVA- OVO- OVA- OVO- OVA- OVO- OVO- OVA- OVA- OVA- OVO-

Vol:.(1234567890)
HCBOUBag HCBOUBag HCBOUBoo HCBOUBoo RBB RBB S+Ada Ada.NC RUS Ada.NC BRF OVO-EE RUS S+Ada BRF OVA-EE
OVO-RBB 8 10 12 9 23 16 17 19 17 17 16 17 14 16 18
OVA-RBB 7 9 11 9 17 13 16 20 16 15 15 17 13 13 16

Scientific Reports |
OVO-S+Ada 10 11 15 15 21 25 23 23 23 20 20 24 23 20 24
OVA-Ada.NC 9 9 12 14 21 22 18 23 21 18 22 22 18 19 23
OVO-RUS 10 10 12 13 21 20 17 19 15 20 18 23 17 20 23
OVO-Ada.NC 10 11 14 15 20 22 18 23 26 21 20 23 19 21 23
OVA-BRF 12 11 15 17 22 24 19 25 24 20 22 23 21 22 25
OVO-EE 12 13 15 15 22 23 18 22 26 22 21 23 19 22 26

(2025) 15:3460 |
OVA-RUS 12 13 14 12 22 23 14 18 22 17 21 19 18 21 24
OVA-S+Ada 14 13 15 14 23 25 21 22 22 21 21 21 22 21 23
OVO-BRF 13 14 16 16 23 26 18 23 24 19 24 22 23 21 27
www.nature.com/scientificreports/

OVA-EE 11 13 13 12 22 24 15 18 22 17 20 19 21 17 20
(f)
OVA- OVO- OVO- OVA- OVO- OVA- OVO- OVA- OVO- OVO- OVA- OVA- OVA- OVO-
HCBOUBag HCBOUBag HCBOUBoo HCBOUBoo RBB RBB S+Ada Ada.NC RUS Ada.NC BRF OVO-EE RUS S+Ada BRF OVA-EE
OVA-HCBOUBag 14 25 24 25 26 21 23 24 23 20 20 20 19 22 21
OVO-HCBOUBag 20 26 23 23 23 22 22 24 22 20 19 19 20 21 19
OVO-HCBOUBoo 6 6 12 18 19 17 19 19 17 13 15 17 17 15 18
OVA-HCBOUBoo 9 9 19 21 20 18 16 20 16 11 17 19 15 16 18
OVO-RBB 5 7 12 9 17 11 13 14 12 9 12 14 11 13 13
OVA-RBB 4 7 11 10 15 11 11 18 12 9 12 14 10 10 14
OVO-S+Ada 9 9 14 12 20 20 18 21 19 15 16 22 19 18 20
OVA-Ada.NC 8 9 11 15 18 20 15 18 19 12 15 20 16 16 19
OVO-RUS 7 7 11 11 17 13 12 15 12 12 12 20 14 13 20
OVO-Ada.NC 8 9 13 15 19 19 15 17 22 12 17 22 17 18 21
OVA-BRF 11 11 17 20 22 22 17 23 21 22 18 21 20 21 21

https://doi.org/10.1038/s41598-024-84786-2
OVO-EE 11 12 15 14 19 19 15 18 21 18 15 23 16 18 25
OVA-RUS 11 12 13 12 17 17 9 13 14 11 13 10 13 14 15
OVA-S+Ada 11 11 14 15 20 21 16 16 17 17 14 16 19 19 20
OVO-BRF 9 10 15 15 18 21 13 18 20 15 15 16 20 14 20
OVA-EE 10 12 12 13 18 17 11 14 13 12 13 10 19 12 15

Table 8.  Pair-wise comparison of methods based on statistical significance across 30 datasets: Number of significant wins in (a) accuracy, (b) averaged-recall, (c) averaged-precision, (d) F1,
(e) G-Mean, and (f) MAUC, metric for the row method over the column method.

16
 Scientific Reports |
(2025) 15:3460 |
www.nature.com/scientificreports/

https://doi.org/10.1038/s41598-024-84786-2
Ada.NC BRF EE HCBOUBoo HCBOUBag RUS RBB S + Ada SUM
OVA OVO OVA OVO OVA OVO OVA OVO OVA OVO OVA OVO OVA OVO OVA OVO OVA OVO
Accuracy 12 10 18 4 5 20 12 14 10 10 7 18 18 9 8 14 90 99
Aver-
aged-pre- 13 11 14 9 8 17 17 12 12 14 10 15 17 11 11 14 102 103
cision
Aver-
aged- 12 11 15 9 4 21 18 11 10 15 9 17 12 16 11 14 91 114
recall
F1 12 12 16 8 5 20 17 12 12 14 9 17 14 14 10 15 95 112
G-mean 7 9 6 8 4 11 17 11 10 13 7 8 7 13 7 9 65 82
MAUC​ 13 11 15 9 5 20 18 11 10 16 10 16 13 15 11 14 95 112
SUM 69 64 84 47 31 109 99 71 64 82 52 91 81 78 58 80

Table 9.  Comparing OVA and OVO for various ensemble techniques and evaluation measures.

17

Vol.:(0123456789)
 www.nature.com/scientificreports/

Fig. 4.  Analyzing the correlation between the imbalance ratio (IR) and performance measures.

Conclusions
Addressing class imbalance is a critical challenge in machine learning, particularly in domains such as healthcare,
fraud detection, and text classification. While binary imbalanced classification has been extensively researched,
multiclass imbalanced classification presents more intricate challenges due to the varying decision boundaries
and complexities inherent in multiple classes. Despite substantial efforts in this area, current solutions often
struggle with issues like overfitting during oversampling and information loss during undersampling. This study
advances the field by introducing the Hybrid Cluster-Based Oversampling and Undersampling (HCBOU) algo-
rithm, which effectively integrates clustering with data-level techniques to overcome these limitations. The
HCBOU algorithm offers a novel approach to multiclass imbalanced classification by employing clustering to
inform the sampling process, ensuring both the preservation of class structure and the generation of meaningful
synthetic instances. This hybrid method improves upon existing techniques by addressing the delicate balance
between reducing redundancy in majority classes and generating relevant data for minority classes, all while
mitigating common issues like overfitting and class distortion. Furthermore, by leveraging one-vs-one (OVO) and
one-vs-all (OVA) decomposition strategies, HCBOU enhances classification performance across diverse datasets,
making it a versatile and powerful tool for real-world applications. The HCBOU algorithm first identifies class
imbalances and applies clustering to divide majority and minority classes into coherent groups. It then performs
undersampling on the majority classes to eliminate redundancy and oversampling on the minority classes to
improve representation. The OVO and OVA decomposition techniques are employed to transform the multiclass
problem into a series of binary tasks, which further enhances the algorithm’s precision. Experimental results
across 30 diverse datasets demonstrate that HCBOU consistently outperforms six state-of-the-art algorithms,
with significant improvements in precision, recall, and F1 scores. These findings confirm that the clustering-based
approach significantly enhances data balance while maintaining the integrity of class relationships. The HCBOU
algorithm holds significant potential for practical applications where minority class prediction is critical, such as
medical diagnosis, fraud detection, and resource management. By improving the balance between minority and
majority classes without compromising data quality, HCBOU enables better generalization in machine learn-
ing models. Its consistent performance across varied datasets underscores its robustness, making it a valuable
contribution to both academic research and industry practices.
While the HCBOU algorithm demonstrates significant improvements in handling multiclass imbalanced
classification, certain aspects could benefit from further refinement. The reliance on clustering methods, while
effective, may increase computational demands in particularly large datasets, although this can be mitigated by
optimizing parameters and using efficient clustering algorithms. Moreover, the choice of clustering technique can
influence performance, but the flexibility of the HCBOU framework allows for adaptation based on the specific
characteristics of the dataset at hand.

Scientific Reports | (2025) 15:3460 | https://doi.org/10.1038/s41598-024-84786-2 18

Vol:.(1234567890)
www.nature.com/scientificreports/

Future research should focus on optimizing HCBOU for large-scale and high-dimensional datasets, poten-
tially by integrating dimensionality reduction techniques to alleviate computational burdens. Experimenting with
alternative clustering methods, such as hierarchical or density-based clustering, could enhance the adaptability of
the algorithm. There is also scope for developing an adaptive version of HCBOU that dynamically adjusts its sam-
pling strategy based on evolving data characteristics. Finally, integrating cost-sensitive learning approaches could
further improve the handling of multiclass imbalances, particularly in time-sensitive or real-time applications.

Data availability
The datasets generated and/or analysed during the current study are available in the [OpenML] repository,
[https://​www.​openml.​org/​search?​type = data&sort = runs&status = active], [Knowledge Extraction Evolution-
ary Learning] repository, [https://​sci2s.​ugr.​es/​keel/​datas​ets.​php] and [UC Irvine Machine Learning] repository,
[https://​archi​ve.​ics.​uci.​edu/​datas​ets].

Code availability
The code used in this study has been made publicly available to enhance reproducibility and verifiability of the
research. It can be accessed at the following GitHub repository: https://​github.​com/​Amir2​7Sale​hi/​HCBOU.

Received: 30 May 2024; Accepted: 27 December 2024

References
1. Zhang, C. et al. Multi-imbalance: An open-source software for multi-class imbalance learning. Knowl. Based Syst. 174, 137–143
(2019).
2. Kim, Y. J., Baik, B. & Cho, S. Detecting financial misstatements with fraud intention using multi-class cost-sensitive learning.
Expert Syst. Appl. 62, 32–43 (2016).
3. Lin, W., Gao, Q., Du, M., Chen, W. & Tong, T. Multiclass diagnosis of stages of Alzheimer’s disease using linear discriminant
analysis scoring for multimodal data. Comput. Biol. Med. 134, 104478 (2021).
4. Haque, R., Islam, N., Tasneem, M. & Das, A. K. Multi-class sentiment classification on Bengali social media comments using
machine learning. Int. J. Cogn. Comput. Eng. 4, 21–35 (2023).
5. Dourado-Filho, L. A. & Calumby, R. T. An experimental assessment of deep convolutional features for plant species recognition.
Eco. Inform. 65, 101411 (2021).
6. Shamrat, F. J. M. et al. High-precision multiclass classification of lung disease through customized MobileNetV2 from chest X-ray
images. Comput. Biol. Med. 155, 106646 (2023).
7. Yang, Y. et al. Data imbalance in cardiac health diagnostics using CECG-GAN. Sci. Rep. 14(1), 14767 (2024).
8. Sánchez-Marqués, R., García, V. & Sánchez, J. S. A data-centric machine learning approach to improve prediction of glioma grades
using low-imbalance TCGA data. Sci. Rep. 14(1), 17195 (2024).
9. Galar, M., Fernandez, A., Barrenechea, E., Bustince, H. & Herrera, F. A review on ensembles for the class imbalance problem:
bagging-, boosting-, and hybrid-based approaches. IEEE Trans. Syst. Man Cybern. Part C (Appl. Rev.) 42(4), 463–484 (2011).
10. Galar, M., Fernández, A., Barrenechea, E., Bustince, H. & Herrera, F. An overview of ensemble methods for binary classifiers in
multi-class problems: Experimental study on one-vs-one and one-vs-all schemes. Pattern Recognit. 44(8), 1761–1776 (2011).
11. Chawla, N. V., Bowyer, K. W., Hall, L. O. & Kegelmeyer, W. P. SMOTE: synthetic minority over-sampling technique. J. Artif. Intell.
Res. 16, 321–357 (2002).
12. Tomek, I. Two modifications of CNN. IEEE Trans. Syst. Man. Cybern. B. SMC-6(11), 769–772 (1976).
13. Gao, X. et al. A multiclass classification using one-versus-all approach with the differential partition sampling ensemble. Eng. Appl.
Artif. Intell. 97, 104034 (2021).
14. Krawczyk, B., Koziarski, M. & Woźniak, M. Radial-based oversampling for multiclass imbalanced data classification. IEEE Trans.
Neural Netw. Learn. Syst. 31(8), 2818–2831 (2019).
15. Li, Q., Song, Y., Zhang, J. & Sheng, V. S. Multiclass imbalanced learning with one-versus-one decomposition and spectral cluster-
ing. Expert Syst. Appl. 147, 113152 (2020).
16. Liu, M., Dong, M. & Jing, C. A modified real-value negative selection detector-based oversampling approach for multiclass imbal-
ance problems. Inf. Sci. 556, 160–176 (2021).
17. Neetha, P., Simran, S., Kainthaje, S. R., Sunilkumar, G., Pushpa, C., Thriveni, J. & Venugopal, K. Borderline-DEMNET for Multi-
Class Alzheimer’s Disease Classification. In 2023 IEEE 5th International Conference on Cybernetics, Cognition and Machine Learning
Applications (ICCCMLA) (2023).
18. Arafat, M. Y., Hoque, S. & Farid, D. M. Cluster-based under-sampling with random forest for multi-class imbalanced classification.
In 2017 11th international conference on software, knowledge, information management and applications (SKIMA) (2017).
19. Dai, Q., Liu, J.-W. & Yang, J.-P. Class-imbalanced positive instances augmentation via three-line hybrid. Knowl. Based Syst. 257,
109902 (2022).
20. Chen, W., Yang, K., Yu, Z. & Zhang, W. Double-kernel based class-specific broad learning system for multiclass imbalance learning.
Knowl. Based Syst. 253, 109535 (2022).
21. Vij, R. & Arora, S. A novel deep transfer learning based computerized diagnostic Systems for Multi-class imbalanced diabetic
retinopathy severity classification. Multimed. Tools Appl. 82(22), 34847–34884 (2023).
22. Ding, S. et al. Kernel based online learning for imbalance multiclass classification. Neurocomputing 277, 139–148 (2018).
23. Ketu, S. & Mishra, P. K. Scalable kernel-based SVM classification algorithm on imbalance air quality data for proficient healthcare.
Complex Intell. Syst. 7(5), 2597–2615 (2021).
24. Li, S. et al. Multi-class imbalance classification based on data distribution and adaptive weights. IEEE Trans. Knowl. Data Eng.
https://​doi.​org/​10.​1109/​TKDE.​2024.​33849​61 (2024).
25. Dai, Q., Liu, J.-W. & Shi, Y.-H. Class-overlap undersampling based on Schur decomposition for Class-imbalance problems. Expert
Syst. Appl. 221, 119735 (2023).
26. Han, M., Guo, H., Li, J. & Wang, W. Global-local information based oversampling for multi-class imbalanced data. Int. J. Mach.
Learn. Cybern. 14(6), 2071–2086 (2023).
27. Tapkan, P., Özbakır, L., Kulluk, S. & Baykasoğlu, A. A cost-sensitive classification algorithm: BEE-Miner. Knowl. Based Syst. 95,
99–113 (2016).
28. Fernández-Baldera, A., Buenaposada, J. M. & Baumela, L. BAdaCost: Multi-class boosting with costs. Pattern Recognit. 79, 467–479
(2018).
29. Iranmehr, A., Masnadi-Shirazi, H. & Vasconcelos, N. Cost-sensitive support vector machines. Neurocomputing 343, 50–64 (2019).

Scientific Reports | (2025) 15:3460 | https://doi.org/10.1038/s41598-024-84786-2 19

Vol.:(0123456789)
 www.nature.com/scientificreports/

30. Liu, W., Zhu, C., Ding, Z., Zhang, H. & Liu, Q. Multiclass imbalanced and concept drift network traffic classification framework
based on online active learning. Eng. Appl. Artif. Intell. 117, 105607 (2023).
31. Yang, J. et al. Deep reinforcement learning for multi-class imbalanced training: applications in healthcare. Mach. Learn. 113(5),
2655–2674 (2024).
32. Mienye, I. D. & Sun, Y. Performance analysis of cost-sensitive learning methods with application to imbalanced medical data.
Inform. Med. Unlocked 25, 100690 (2021).
33. Liu, X. Y., Wu, J. & Zhou, Z. H. Exploratory undersampling for class-imbalance learning. IEEE Trans. Syst. Man Cybern. Part B
(Cybern.) 39, 2539–2550 (2008).
34. Seiffert, C., Khoshgoftaar, T. M., Van Hulse, J. & Napolitano, A. RUSBoost: A hybrid approach to alleviating class imbalance. IEEE
Tran. Syst. Man Cybern.-Part A: Syst. Humans 40(1), 185–197 (2009).
35. Grina, F., Elouedi, Z. & Lefevre, E. Re-sampling of multi-class imbalanced data using belief function theory and ensemble learning.
Int. J. Approx. Reason. 156, 1–15 (2023).
36. Wang, S. & Yao, X. Diversity analysis on imbalanced data sets by using ensemble models. In 2009 IEEE symposium on computational
intelligence and data mining (2009).
37. Hido, S., Kashima, H. & Takahashi, Y. Roughly balanced bagging for imbalanced data. Stat. Anal. Data Min. ASA Data Sci. J. 2(5–6),
412–426 (2009).
38. Wang, S., Chen, H. & Yao, X. Negative correlation learning for classification ensembles. In The 2010 international joint conference
on neural networks (IJCNN) (2010).
39. Zhang, Z., Krawczyk, B., Garcia, S., Rosales-Pérez, A. & Herrera, F. Empowering one-vs-one decomposition with ensemble learning
for multi-class imbalanced data. Knowl. Based Syst. 106, 251–263 (2016).
40. Chen, C., Liaw, A. & Breiman, L. Using random forest to learn imbalanced data. Univ. California, Berkeley 110(1–12), 24 (2004).
41. Rodríguez, J. J., Diez-Pastor, J.-F., Arnaiz-Gonzalez, A. & Kuncheva, L. I. Random balance ensembles for multiclass imbalance
learning. Knowl. Based Syst. 193, 105434 (2020).
42. Dai, Q., Wang, L.-H., Xu, K.-L., Du, T. & Chen, L.-F. Class-overlap detection based on heterogeneous clustering ensemble for
multi-class imbalance problem. Expert Syst. Appl. 255, 124558 (2024).
43. Rousseeuw, P. J. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. J. Comput. Appl. Math. 20, 53–65
(1987).
44. LemaÃŽtre, G., Nogueira, F. & Aridas, C. K. Imbalanced-learn: A python toolbox to tackle the curse of imbalanced datasets in
machine learning. J. Mach. Learn. Res. 18(17), 1–5 (2017).
45. Sokolova, M. & Lapalme, G. A systematic analysis of performance measures for classification tasks. Inf. Process. Manage. 45(4),
427–437 (2009).
46. Hossin, M. & Sulaiman, M. N. A review on evaluation metrics for data classification evaluations. Int. J. Data Min. Knowl. Manag.
Process 5(2), 1 (2015).
47. Kubat, M., Holte, R. C. & Matwin, S. Machine learning for the detection of oil spills in satellite radar images. Mach. Learn. 30,
195–215 (1998).
48. Hand, D. J. & Till, R. J. A simple generalisation of the area under the ROC curve for multiple class classification problems. Mach.
Learn. 45, 171–186 (2001).
49. Wilcoxon, F. Individual comparisons by ranking methods. Breakthr. Stat.: Methodol. Istrib. https://d ​ oi.o
​ rg/1​ 0.2​ 307/3​ 00196​ 8 (1992).

Author contributions
Amirreza Salehi: Conceptualization, Methodology, Data Analysis, Writing—Original Draft. Majid Khedmati:
Conceptualization, Supervision, Writing—Review & Editing.

Funding
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-
profit sectors.

Declarations

Competing interests
The authors declare no competing interests.

Additional information
Correspondence and requests for materials should be addressed to M.K.
Reprints and permissions information is available at www.nature.com/reprints.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which
permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give
appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and
indicate if changes were made. The images or other third party material in this article are included in the article’s
Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included
in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or
exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy
of this licence, visit http://creativecommons.org/licenses/by/4.0/.

© The Author(s) 2025

Scientific Reports | (2025) 15:3460 | https://doi.org/10.1038/s41598-024-84786-2 20

Vol:.(1234567890)