Resampling approaches to handle class imbalance: a review from a data perspective

The review of the literature reveals four main determinants of classification complexity, all of which exhibit some form of interdependence, which are the existence of small disjuncts, noisy data, data-shifts and class overlap [4, 5, 16]. Small disjuncts correspond to the existence of different sub concepts or clusters for the same class, usually with a small number of instances. Noisy data refers to inconsistencies that negatively affect the data, being further subdivided into feature noise or class noise, the latter having the most impact on the predictive ability of the classifier [4]. As for data-shifts, it occurs when the assumption that the distributions of the training data and the test data are similar is violated, limiting the generalization capabilities of the learned models [4]. Finally, class overlap is the phenomenon most thoroughly described in the literature, corresponding to the existence of regions populated by samples of different classes [16], impairing the clear definition of a decision boundary capable of distinguishing between classes, without inducing overfitting. Although this phenomenon is conceptually intuitive, its quantification remains an open question within the scientific community, since there are no standardized guidelines for this matter [5, 16]. This fact arises on account of class overlap being composed by several different forms of complexity [16], making it difficult to define a metric capable of encapsulating and describing its nature as a whole. In an attempt to illustrate this, we present three distinct imbalanced scenarios in Fig. 1, each showcasing a unique form of class overlap, subsequently leading to diverse classification complexities. Within Fig. 1a, the overlapped region demonstrates an imbalance ratio akin to the global imbalance of the classification problem, whereas Fig. 1b presents an overlapped region with a minority sample density four times greater than the remainder of the class. Meanwhile, the overlapped region in Fig. 1c also presents a greater sample density, albeit scattered into numerous subclusters. These scenarios underscore how diverse data irregularities, such as local imbalance in Fig. 1a, b, or the presence of subclusters in Fig. 1c, as well as other data difficulty factors like highly non-linear boundaries or high data dimensionality, give rise to various types of class overlap, each with unique impacts on classifier performance [5]. Hence, we provide a brief summary describing how the various manifestations of class overlap affect classifier performance within imbalanced domains. That being said, for an in-depth discourse on this topic, we direct the reader to [5].

The synergistic effect between class imbalance and class overlap depends on the local characteristics of the data in the overlap region (local imbalance and data typology), general characteristics of the domain (class decomposition, dimensionality, decision boundary complexity) and, finally, the nature of the classifier at hand [5]. As for the first topic previously mentioned, the literature has proven that, if the overlap region presents a local distribution of classes similar to the global imbalance of the data (Fig. 1a), classifiers with a global nature (e.g., Linear Discriminant Analysis) will be much more biased towards the majority class. Conversely, classifiers with a more local nature (e.g., k-NN) are shown to exhibit superior ability in this experimental setting [5]. On the other hand, data typology, which refers to the categorization of samples based on their neighborhood, has been shown to be a major factor in increasing classification complexity. The most commonly adopted protocol for data topology analysis, formulated by Napierala et al., requires examination of the five nearest neighbors to each instance [9]. Within this schema, instances with at most one nearest neighbor from a differing class are designated as "safe". Instances with less than three opposing class neighbors fall under the "borderline" category. Samples with only one same-class neighbor are deemed "rare" and those encircled purely by contrasting class samples are identified as "outliers" [15]. In this context, the existence of a significant number of borderline samples, characterized by their disposition at the boundary between classes, limits the performance of classifiers. Shifting the focus to the general characteristics of the data, domains with a more complex structure, evidenced through the existence of highly non-linear decision boundaries, or with the existence of multiple sub concepts culminate in performance decreases, especially in tree and rule-based classifiers [5]. Lastly, we highlight that data dimensionality was shown to have a relatively small impact on classification performance, although very few studies have been conducted in this regard [5].

Complexity metrics are measures designed to quantify intrinsic characteristics of the data which imply some difficulty to the classification problem, for instance its topology or local imbalance [5]. The most widely reported approach in the literature of complexity evaluation, designated Maximum Fisher's discriminant ratio (\(F1\)) [5], intents to assess the projection of the data that maximises separability between classes, as described by the following formulation:with \({f}_{j}\) referring to each feature, where \(j=1,\dots , m\), \({\mu }_{{f}_{j}}^{c0}\) and \({\sigma }_{{f}_{j}}^{c0}\) indicating the mean and standard deviation of \({f}_{j}\) in the majority class and \({\mu }_{{f}_{j}}^{c1}\) and \({\mu }_{{f}_{j}}^{c1}\) pertaining to the mean and standard deviation of \({f}_{j}\) in the minority class. Although this metric provides a more global analysis about the overall distribution of the data, it has limited ability to assess the local properties of the data, which is extremely relevant when analysing the boundaries between classes, as they usually correspond to the most problematic regions for the classification process [5]. In this case, instance overlap features are proposed based on the premise that not all examples in the data present the same degree of classification complexity (concept typically referred to as instance hardness), thus triggering the need to introduce more granularity in the analysis of the complexity of the classification task [5]. In this context, one resorts, for instance, to the error associated to a k-NN classifier with 1 neighbour in synthetic data generated through interpolation based on the original dataset, which is commonly referred to as the N4 complexity metric. Immediately, the difficulty and heterogeneity related to the quantification of this phenomenon is evident. To further this discussion, the F1 and N4 metrics were applied to the artificial data sets presented in Fig. 1a, b. Given that both datasets possess similar means and standard deviations, differing only in the overlap region, the calculated F1 values were nearly indistinguishable, recording 0.816 and 0.828 for Fig. 1a, b, respectively. However, when the N4 metrics were employed, a significant divergence was noted, with the complexity of the Fig. 1b dataset (0.079) being more than double that observed for the Fig. 1a dataset (0.033). This increase in complexity is explicable by the greater incidence of misclassification of minority class samples within the overlap region.

These results reveal how, depending on the utilized metric and on the nature of the complexity evaluated, completely different evaluations of the complexity of the classification task are obtained, since the presented scenarios exhibit almost identical \(F1\) values and, simultaneously, quite divergent 1-NN error values. This problem stands out as one of the main focuses of research in this area, where new indicators of class overlap, which consider the different types of complexity, should be developed [5], thereby providing a more through description of the characteristics of the data. In the scope of resampling approaches, the lack of a methodology capable of holistically evaluate classification complexity and identify regions of interest can bottleneck the capabilities of these approaches, as will be discussed in subsequent sections. Additionally, the multitude of class overlap representations has hindered the definition of a standardised taxonomy that makes it feasible to categorise the different developed metrics, thus enhancing a systematic analysis and comparison of the approaches reported in the literature [5, 16]. The most frequently employed metric categorization was developed by Ho and Basu, assuming the division of complexity metrics into: \((i)\) measures of overlap between individual features, \((ii)\) measures of separability between classes and \((iii)\) measures of geometry, topology, and density of manifolds [8]. Similarly, new taxonomies have been proposed over the years to accompany the new developments in the area, among which stands out the work recently developed by M. Santos et al. establishing the classification of overlap-based complexity metrics into four distinct categories [5]:

To conclude the discussion on class overlap, we emphasise the fact that most of the presented metrics do not acknowledge the presence of class imbalance, often being bias towards the majority class. Returning to the example in Fig. 1 and recalling that the F1 indicator depends exclusively on the mean and standard deviation of the features, it is agnostic to the proportion between the minority class and the majority class. Similarly, applying the 1-NN metric to a highly imbalanced dataset and hypothesizing a worst-case scenario of classification complexity where every minority sample is misclassified by the 1-NN, the resulting total error will nonetheless be minor. Therefore, in light of the fact that class imbalance significantly contributes to the complexity of the classification task, the majority of conventional complexity metrics are ill-equipped to provide insights in such domains.

Considering the relevance of jointly evaluating imbalance and data irregularities in scope of resampling approaches, coupled with the limitations of traditional complexity metrics, several scientific works suggested modifications to the established metrics or have introduced entirely new metrics, which are succinctly compiled in Table 1. These are usually proven to be more accurate than their traditional counterparts, by yielding a higher correlation with classification performance, usually evaluated by metrics sensitive to class imbalance, such as G-mean or F1-score [14].

By analysing Table 2, it becomes evident that the most commonly employed strategy to circumvent the metrics´ bias towards the majority class is to compute the metrics individually for the different classes. This is understandable, since most of the original metrics normalize the data according to the global number of samples, limiting the impact of the minority class. It is noteworthy that many resampling approaches resort to metrics similar to those depicted in Table 1 to adaptively tailor their behaviour to each classification problem, as will be further discussed in “Resampling approaches and complexity assessm”.

This section aims to outline the primary approaches for assessing the performance of learning models in imbalanced problems. It should be emphasized that these performance metrics differ from the complexity metrics previously addressed, as their goal is to provide a holistic perspective on the model's efficacy based on predefined user preferences. Although these can also be interpreted as a global estimate of classification complexity, they fail to offer detailed insights into the data irregularities underlying such classification complexity. That being said, performance metrics are occasionally applied in complexity assessment, such as in the L1 or N3 metric [14].

Traditional performance metrics, such as accuracy and its complement, the error rate, encounter a significant limitation in imbalanced datasets, as the metric’s value is largely driven by the most frequent class [1, 6]. This leads to biased evaluations favouring the majority class in imbalanced domains. Since, in such classification tasks, correct identification of the minority class is often of higher importance, these metrics are inadequate for assessing model performance [1, 20]. Note that, selecting a performance metric that aligns with the specific requirements of the classification problem is crucial, as it profoundly influences model selection, parameter tuning, and other aspects of the modelling pipeline [1].

To introduce the metrics commonly employed in imbalanced domains, we begin by considering a binary classification problem with one positive and one negative class. The classifier’s performance is summarized using a confusion matrix, depicted in Table 2. This matrix categorizes the samples as correctly classified—True Positives (TP) for the positive class and True Negatives (TN) for the negative class—or misclassified—False Positives (FP) and False Negatives (FN), depending on the true class [21]. Based on this foundation, we proceed to present a series of metrics tailored for imbalanced domains, which can be applied according to user-defined preferences on desired model behaviour.

The preceding formulations demonstrate that different metrics evaluate distinct aspects of a classifier’s performance. For instance, recall quantifies the proportion of positive samples correctly identified, while specificity measures the proportion of negative samples correctly classified [1]. In many experimental contexts, improving recall often leads to a decline in specificity, underscoring the importance of simultaneously tracking both metrics to strike a balance that aligns with the desired behaviour of the model [1, 20]. A similar interplay exists between other metrics, such as the true positive rate and true negative rate or precision and recall. To facilitate the analysis of such trade-offs, integrated metrics like the Fβ score and the geometric mean are adopted, which are formalized below.

The Fβ score serves as the harmonic mean between precision and recall, with the parameter β dictating the relative importance of these two measures. Values of β greater than 1 emphasize recall over precision, thereby imposing a higher penalty on false negatives than false positives. In contrast, β values below 1 place greater weight on precision, reducing the impact of false negatives [21]. A commonly used configuration is \(\beta = 1\), known as the F1-score, which assigns equal weight to precision and recall. The G-mean, on the other hand, is calculated as the geometric mean of the accuracies for both classes, offering a balanced assessment of model performance [7]. Additionally, it is worth noting other similar metrics have been proposed in the literature, such as dominance index, balanced accuracy, and mean class-weighted accuracy [1].

Another critical aspect in this context is the extension of these evaluation metrics to multiclass scenarios. This is commonly achieved by calculating metrics independently for each class and subsequently aggregating these values using one of two approaches: micro-averaging or macro-averaging. Micro-averaging averaging applies weights proportional to the frequency of each class, whereas macro-averaging gives equal importance to all classes, making it particularly advantageous in imbalanced classification settings [1]. Readers seeking a detailed mathematical formalization of metric computation in these contexts are encouraged to consult [1].

Beyond the scalar metrics previously examined, graph-based metrics are essential for assessing classifier performance, with the Receiver Operating Characteristic (ROC) curve being a preeminent example [1]. These metrics are invaluable when the problem's decision-making preferences are not predefined, as they comprehensively report performance across the full spectrum of operating conditions [1]. The ROC curve quantifies performance by evaluating classifier outcomes at different decision thresholds, which determine the probability value at which samples are assigned to the positive class (the default threshold in most classifiers is typically set at 0.5) [20, 21]. Each point on the curve represents a true positive rate and a false positive rate corresponding to a particular threshold. A significant limitation of the ROC curve lies in its inability to facilitate straightforward comparisons between classifiers. To overcome this drawback, the Area Under the ROC Curve (AUC-ROC) is widely adopted, providing a unified measure that encapsulates overall classifier performance across various operating conditions, simplifying comparative analysis [1]. Additionally, other graph-based metrics, including Precision-Recall curves, Prob AUC, Scored AUC, and Brier curves, have emerged to address certain limitations inherent to the ROC-based methodology.

In the context of addressing class imbalance using resampling techniques, SMOTE has gained widespread recognition as one of the most extensively utilized oversampling techniques [23]. It introduced a paradigm shift by pioneering the creation of synthetic samples, distinguishing itself from the prevailing practice of replicating existing samples, as exemplified by Random Oversampling (ROS) [24]. The shift towards synthetic sample generation stemmed from the limitations of ROS, particularly its tendency to induce model overfitting by promoting the creation of very specific decision regions through the iterative duplication of identical samples [25]. To address this challenge, SMOTE introduces a strategy wherein the generation of new samples is achieved through interpolation. This technique leverages the proximity of samples belonging to the minority class, enabling a more accurate and comprehensive modelling of the minority class. Consequently, the model's ability to generalize improves considerably [24, 25].

Formally, SMOTE method begins by selecting the amount of oversampling to be performed, \(N\), which is typically chosen to obtain a 1:1 ratio between the majority and minority samples [24]. Afterwards, an instance of the minority class, \({x}_{i}\), is randomly selected, from which the k-nearest neighbours pertaining to the same class, \(NN({x}_{i})\), are determined, where \(k\) typically assumes the value of 5 [24]. Lastly, \(N\) of the \(k\)-nearest neighbours are selected to generate new samples, \({x}_{new}\), based on the following expression:

With \(rand\left(\text{0,1}\right)\) indicating a randomly selected number between 0 and 1. By repeatedly executing this process, the desired number of samples can be generated to attain a balanced dataset. Adjusting class frequencies to equality prevents the misclassification cost of the majority class, driven by its larger sample size, from dominating the classifier training process. This approach enhances the likelihood of correctly classifying minority class instances. Empirically, resampling has been shown to lead to statistically significant gains in classifier performance, including improvements in recall and F1-score [26].

However, despite its widespread adoption, several limitations arise in its application. Notably, these limitations stem from the inherent randomness associated with sample selection for oversampling, where samples are chosen with uniform probability. Consequently, regions with higher density of minority samples have higher probability of being further inflated and vice-versa, thus depicting certain difficulties of this method in addressing within-class imbalance and the presence of small disjuncts. Moreover, the utilization of SMOTE has the potential to magnify the impact of noise in the data by oversampling noisy samples, ultimately leading to diminished performance or overfitting. Lastly, it is worth noting that the structural characteristics of the majority class are not considered during the generation of new samples, which may result in class overlap [25]. To mitigate such issues, several SMOTE variants have been developed and are readily accessible through popular machine learning libraries such as scikit-learn [27].

In 2005, Han et al. proposed the Borderline-SMOTE method, which is based on the premise that the samples at the boundaries between classes are the most determinant in defining an adequate decision boundary, i.e., that maximises the distinction between the different classes [28]. This being the case, the algorithm initially foresees the categorization of samples as safe, danger or noise, depending on the proportion of majority samples within the nearest neighbours of the minority samples. Note that samples are considered dangerous if half of the nearest neighbours correspond to the majority class, unless all nearest neighbours originate from the contrary class, in which case they are identified as noise. Once the danger samples are determined, which are typically arranged along the border between classes, the artificial sample generation procedure is initiated based on Eq. (11) [28]. Under the same underlying hypothesis, Hguyen et al. introduced a novel technique known as SVMSMOTE [29]. This method performs oversampling of support vectors belonging to the minority class, which are obtained through the training of a SVM on the original dataset. In SVMSMOTE, the generation of new samples encompasses two distinct strategies: extrapolation and interpolation. When the number of minority neighbours exceeds that of the majority samples, extrapolation is employed, thereby inducing an expansion of the minority class region. Conversely, in scenarios where the inverse holds true, interpolation is applied to consolidate the minority class area. Importantly, in contrast to SMOTE, the selection of samples in SVMSMOTE occurs sequentially, by the evaluation of their proximity to the sample under consideration, \({x}_{i}\) [29].

In 2008, He et al. suggested the ADASYN method which diverged from SMOTE by introducing a novel strategy for adaptive generation of synthetic samples [30]. This strategy considers the density distribution of the minority class, which is evaluated based on the number of samples from the majority class present within the \(k\)-nearest-neighbours of each sample. By incorporating this density information, regions characterized by a scarcity of minority samples are prioritized, resulting in a higher number of synthetic samples being generated in such areas through SMOTE´s procedure. Consequently, ADASYN provides enhanced representation for challenging instances that are more difficult to learn, thus contributing to the effective handling of class imbalance [30].

It is important to note that the methods described above represent only a subset of the extensive research conducted on data-driven approaches to handle class imbalance. Numerous other algorithms, such as Safe-Level SMOTE [31], MWMOTE [32], DBSMOTE [33], and more, have been proposed in the literature. To gain a comprehensive understanding of the advancements in this area, we direct the reader to the article [25], which offers a systematic review of various SMOTE variants up until the year 2018. In the aforementioned research, the authors proposed a framework to categorize the different changes performed to the original SMOTE algorithm. In this work, we adopt a modified version of this taxonomy for oversampling techniques as outlined in Fig. 2. The methods are first classified into interpolation-based or non-interpolation-based approaches based on methods’ applied principle for sample generation. Then, within the interpolation-based methods, six of the seven categories of SMOTE modifications proposed in [25] are adopted, excluding the "integration with undersampling" category, which is viewed as an hybrid sampling technique in this study. As a result, interpolation-based methods can fall under multiple categories simultaneously if they incorporate several modifications to the original SMOTE. Given that the majority of recent oversampling techniques are interpolation-based, we believe this revised taxonomy effectively encompasses the existing body of literature. It is important to note that this section focuses solely on interpolation-based methods, while “Non-interpolation based oversampling approaches” addresses recently introduced non-interpolation-based approaches.

Based on the presented taxonomy, we intend to continue the efforts reported in [25], and conduct a thorough review of the new SMOTE-based approaches reported in the literature from 2019 to date. To ensure clarity and coherence, we present our findings in the form of a table, which makes straightforward a concise summary of taxonomy, methodology and conclusion from the various reported works.

The evolutionary trajectory of undersampling approaches has mirrored that of oversampling approaches, since the first proposed method was Random Undersampling (RUS), which involved the random removal of instances from the majority class. By doing so, RUS imposes an identical significance to both classes during the optimization process of classifiers. However, the application of RUS carries a substantial risk of information loss, thereby adversely impacting the classifier's performance [1]. Consequently, informative undersampling techniques have emerged, incorporating data characteristics to selectively eliminate samples to enhance the visibility of the minority class. Although numerous undersampling approaches have been proposed in the literature, this section will be focused on most used techniques and available in open-source repositories.

In response to the intrinsic loss of information associated with random sample removal, one of the initial methods introduced was the Condensed Nearest Neighbours (CNN) technique [65]. This approach firstly foresees the separation of minority class samples into a dedicated set, denoted as \(C\), while assigning the majority class samples to set \(S\). Subsequently, majority samples are added to set \(C\) one by one and a 1-NN classifier is employed to determine whether the added sample is correctly classified. In the case the sample is misclassified, it is permanently integrated into set \(C\). This iterative process is repeated until no further samples can be added and the obtained set \(C\) will correspond to the new balanced dataset. The primary objective of CNN is to derive a representative subset of the data distribution that exhibits performance on par with that obtained using the complete dataset, this concept being termed minimal consistent subset [65]. However, it is important to note that this algorithm has some vulnerability to noisy samples, since these will often be misclassified and thus included in set \(C\). To address this limitation, Kubat et al. proposed the One-Sided Selection (OSS) method, where after applying CNN, the majority samples that integrate Tomek-Links are removed [75]. Tomek-Links are pairs of samples with distinct classes that are nearest neighbours to each other and, as such, normally correspond to instances located at the border between classes [76]. Another prominent approach, which shares a similar concept with the OSS technique, where all minority class samples are retained in the final balanced dataset, is the Neighbourhood Cleaning Rule (NCR) [77]. The NCR technique leverages the Edited-Nearest Neighbours (ENN) method to identify outliers within the majority class. ENN is an undersampling technique that removes samples whose labels differ from the majority of their k-nearest neighbours [77]. Returning to NCR, a subsequent step involves employing a 3-Nearest Neighbour (3-NN) classifier on the minority class instances. If misclassifications occur, the nearest neighbours from the majority class are eliminated. It is important to note that this description has been simplified to the context of binary classification, whereas the original algorithm is designed to handle multiclass problems, as thoroughly described in [77].

In 2003, Zhang et al. suggested 3 different undersampling methods based on neighbourhood analysis, namely NearMiss-1, NearMiss-2 and NearMiss-3. NearMiss-1 provides for removing majority samples whose average distance to the \(k\)-nearest neighbours of the minority class is the smallest (original article used \(k=3\)). In contrast to NearMiss-1, NearMiss-2 assumes the removal of majority samples whose distance from the farthest \(k\)-nearest neighbours are the smallest. Finally, NearMiss-3 involves two consecutive steps, where first the majority \(M\) nearest neighbors of each minority class are determined. Then, within the defined \(M\) nearest neighbours, those whose average distance to the 3 minority nearest neighbours is closest are eliminated [78].

Lastly, we draw attention to the methodology introduced by Smith et al., which is centred around the notion of instance hardness (\(IH\)) [79]. IH represents the probability of a sample being misclassified and serves as a metric for assessing the classification complexity of a particular instance, as introduced in "Classification complexity in imbalanced domains". Formally, the concept of instance hardness can be derived from Bayes theorem and assumes the following formulation exposed in Eq. (12).where \({x}_{i}\) and \({y}_{i}\) represent the data and corresponding labels and \(h\) refers to the learning algorithm, thus adhering to the notation outlined in "Class imbalance: formal definition". As one can observe, the computation of instance hardness is dependent on the chosen classifier, however, the authors suggest adopting a weighted average of the instance hardness obtained with several learning models, \(\mathcal{L}\), as exposed in Eq. (13), in order to attain a general assessment of instance hardness [79].

From a practical standpoint, the estimation of the conditional probability \(p\left({y}_{i}|{x}_{i},h\right)\) relies on the scores derived from the classifier \({h}_{j}\). As for the selected classifiers, clustering technique based on Classifier Output Difference was employed [80], which enables the attainment of a diverse set of learning algorithms, thus generating a representative value of \(\text{IH}\). In order to effectively apply the quantification of instance hardness for informative undersampling, the authors put forth an approach involving the utilization of a threshold. Specifically, samples characterized by an \(IH\) value surpassing a predetermined threshold (0.7 exhibited the best results in the conducted research) are designated for removal [79].

To complement the discussion of undersampling methods, it is important to highlight the study conducted by Wainer et al. [26], which presented a comparative analysis of all the aforementioned methods across 58 imbalanced datasets from the KEEL repository. Experimentally, a threefold repetition of a stratified 20% holdout procedure was used, with a nested threefold cross-validation to select the optimal hyperparameters for each undersampling method, where the classifier was also treated as a hyperparameter. After determining the performance of each approach using 8 distinct performance metrics, a statistical analysis was carried out to determine if there were significant differences between the methods. More specifically, pairwise Wilcoxon tests were applied to the ranks of each method, with a multiple comparison adjustment of the p-value. The results showed that, for the F1-score, NCR had the lowest mean rank, although it did not exhibit significant differences compared to Near-Miss, CNN, and IH.

After reviewing some of the most relevant approaches within undersampling techniques to address class imbalance, an overview of the literature regarding the most recently reported approaches is presented next, as shown in Table 4. Furthermore, in order to enhance the clarity and systematic analysis of the undersampling approaches, we adopt a taxonomy framework proposed in [1]. In this taxonomy, resampling approaches are initially categorized into three main groups: stratified approaches, synthetic data generation methods, and hybrid approaches. Among the stratified approaches, which primarily involve undersampling techniques, further subdivision is performed, resulting in six distinct categories: distance-based, recognition-based, evolutionary-based, cluster-based methods, data cleaning based and random under/over sampling. A schematic representation of the adopted taxonomy is presented in Fig. 3.

Distance-based methods leverage inter-class or intra-class distances to select pertinent samples. Recognition-based approaches operate by utilizing information from a single class to perform informed resampling, deviating from the traditional discrimination-based inductive learning paradigm [1]. These approaches excel in extreme imbalance scenarios where the available data is insufficient to adequately approximate the underlying data distribution. Examples of recognition-based approaches include Autoencoders and one-class SVMs. Cluster-based resampling methods employ clustering algorithms to model either one or both classes, whereas evolutionary methodologies utilize evolutionary algorithms to optimize a predefined fitness function, ultimately generating a resampled dataset. Cleaning-based methods aim to remove outliers and address instances with overlapping characteristics, as these factors significantly impact the performance of learning models [5]. Lastly, random-based approaches randomly select samples for removal [1]. It is worth noting that certain proposed methods encompass multiple steps, such as combining data cleaning and cluster-based undersampling, resulting in approaches that can simultaneously belong to multiple categories.

As mentioned earlier, the utilization of oversampling methods in data resampling introduces the potential risk of amplifying noise and inducing class overlap, thereby increasing the likelihood of overfitting. Conversely, undersampling methods face the challenge of removing informative samples necessary for accurate model training, resulting in underfitting [95, 96]. To overcome the limitations associated with both approaches, the emergence of hybrid resampling methods has provided a solution by simultaneously employing oversampling and undersampling methods, either in an alternate or sequential manner [45, 96]. Another motivation behind the development of hybrid approaches stems from previously conducted analysis in which no single method seems to guarantee the most competitive performance in all classification contexts. Given the observation that oversampling methods do not consistently outperform undersampling methods and that both can be advantageous under specific circumstances, their integration represents a logical alternative for the creation of more robust and consistent resampling methodologies [95]. Typically, the resampling process commences with the implementation of oversampling techniques, followed by the introduction of undersampling measures aimed at mitigating the occurrence of class overlap through the removal of majority samples, ultimately culminating in the construction of a balanced dataset. In essence, the role of undersampling is analogous to the data cleaning procedures employed post-oversampling with SMOTE, with the distinction that data cleaning operates on the generated minority samples instead of eliminating majority samples as observed in undersampling.

One of the initial hybrid methods that emerged in the field of data resampling was the combination of SMOTE with Tomek links [45] and SMOTE with ENN [58]. Regarding the first method, it foresees the application of SMOTE to generate synthetic samples, followed by the determination of Tomek links present in the balanced dataset. Finally, the majority samples that integrate Tomek links are eliminated [45]. On the other hand, the SMOTE + ENN approach begins with the application of SMOTE to augment the minority class, and subsequently employs the ENN algorithm to address misclassified samples, irrespective of their class (majority or minority). As such, the data cleaning process conducted by SMOTE + Tomek links is more comprehensive [58]. Over the years, a wide range of hybrid resampling methods have been proposed. In line with the preceding sections, we aim to analyse the recent trends in the literature pertaining to this topic, as described in Table 4.

Deep learning is a subfield of machine learning that relies on models built with multi-layered architectures, enabling the learning of hierarchical data representations [111]. This mimics, to some extent, the structure and functionality of the human brain [111, 112]. These layers are typically tailored to the nature of the data being analysed, for instance, 2D convolutional layers are often used for analysing spatial patterns in images, while transformer or Long Short-Term Memory (LSTM) blocks are suited for processing sequential data [111]. One key differentiator of DL from traditional ML approaches is their ability to bypass the need for many handcrafted preprocessing steps such as feature selection. Instead, DL models can learn representations directly from raw data, providing an unprecedented level of flexibility and ability to capture complex patterns in the data [111]. However, this capability comes at the cost of increased model complexity, imposing substantial computational resources. That being said, the evolution of specialized hardware and software for highly efficient and parallelized model training, alongside the increased availability of large-scale datasets, has alleviated these limitations, enabling the development of DL systems that consistently outperform conventional (shallow) models [111, 112]. For a comprehensive examination of state-of-the-art deep learning frameworks, we direct the reader to [111].

Nevertheless, DL models, much like traditional ML methods, struggle with challenges associated with imbalanced data distributions. This issue is compounded by factors such as data drift, high number of classes, and the relationship between imbalanced data and the extracted embeddings [113]. To address these limitations, numerous approaches have been proposed, spanning both data-driven and algorithmic-driven solutions. Within the scope of this article, we concentrate on interpolation-based data-driven methods, such as those elaborated upon in "Oversampling", "Undersampling approaches" and "Hybrid approaches".

In [114], the authors proposed a network detection system (NDS) leveraging a hybrid sampling technique combined with a one-dimensional convolutional neural network (1D-CNN). The study utilized two open-source datasets: UNSW-NB15, comprising 2.54 million samples with 87.35% representing normal traffic and 12.65% attack traffic, and CICIDS2017, consisting of approximately 2.8 million samples, of which 80.3% correspond to normal traffic and the remaining data span 14 distinct attack categories [114]. Following preprocessing steps such as one-hot encoding, feature reduction, data standardization, and feature selection, resampling was applied using a specialized method. In this approach, the target sample size per class \({I}_{resample}\) was defined as the total number of samples divided by the number of classes. For classes with fewer samples than \({I}_{resample}\), SMOTE was employed, while classes exceeding this threshold underwent Gaussian Mixture Model (GMM)-based clustering for sample extraction [114]. Experimentally, the proposed method consistently achieved higher F1-scores in both binary and multiclass classification tasks compared to approaches combining Random Forest (RF) and Multi-Layer Perceptron (MLP) with traditional resampling techniques such as ROS, SMOTE, ADASYN, RUS + SMOTE, and K-means + SMOTE. Additionally, the approach outperformed other state-of-the-art methodologies [114]. In the context of cybersecurity, Abdelkhalek et al. aimed to develop a network detection system (NDS) based on the open-source dataset NSL-KDD, which includes normal traffic data and four types of attacks: Denial of Service (DoS), Probing, Remote-to-Local (R2L), and User-to-Root (U2R) attacks [115]. To address the class imbalance present in the dataset, the authors employed a combination of ADASYN and Tomek Links. This resampling approach was evaluated using various models: Convolutional Neural Network (CNN), Multi-Layer Perceptron (MLP), Deep Neural Network (DNN), and a hybrid CNN-BLSTM architecture. Experimentally, it was observed that in binary classification scenarios, the CNN model combined with ADASYN and Tomek Links achieved the most competitive F1-score (84%) [115]. Conversely, for multiclass classification, the MLP model coupled solely with oversampling yielded the best performance (87%). Additionally, independent of the model used, the application of oversampling consistently resulted in significant performance improvements [115]. In [116], the authors also developed a method addressing class imbalance through the combination of random undersampling of the majority class (normal traffic) and SMOTE applied to two minority classes (Web Attacks and Infiltration Attacks). Classification was performed using a model comprising two LSTM layers and three fully connected layers [116]. The dataset employed in this study was CSE-CIC-IDS2018, which contains eight classes, seven of which correspond to distinct network attack types. Experimental results demonstrated that the adopted architecture outperformed traditional machine learning algorithms such as Decision Trees, Random Forests, k-Nearest Neighbours, and Support Vector Machines [116]. In the context of fraudulent transaction detection, the authors in [117] utilized SMOTE-ENN to generate synthetic samples due to the severe class imbalance in the European credit card dataset, where only 0.172% of the samples represented fraudulent transactions. Following this resampling step, an ensemble model was trained, consisting of a GRU network and an LSTM network, whose predictions were combined via a MLP acting as a meta-classifier [117]. Experimental results demonstrated that incorporating resampling into this approach enhanced sensitivity, specificity, and AUC by 10.5%, 3.2%, and 8.7%, respectively. Furthermore, the proposed method was tested on another dataset (Taiwan dataset), where similar performance improvements were also observed, along with superiority over other machine learning algorithms such as AdaBoost, Random Forests (RF), MLPs, LSTMs, and GRU networks [117]. In the domain of emotion detection using EEG signals, motivated by the rising prevalence of health issues related to stress and anxiety, the authors in [118] developed a methodology to evaluate degrees of arousal and valence, categorized as positive, medium, or negative, based on physiological signals. Data was sourced from the DEAP database, which comprises EEG recordings from 32 participants collected through 32 electrodes. After preprocessing, spectral power features from five frequency bands were extracted using the Short-Time Fourier Transform applied to 2-s sliding windows [118]. Subsequently, Borderline SMOTE was applied for oversampling, and the resulting augmented data were used to train a 1D-CNN. Results showed an increase in F1-score from 92.04% to 97.54% for valence classification and from 92.75% to 97.84% for arousal classification, following the introduction of resampling [118].

In the field of medical applications, the authors in [119] assessed three network architectures, CNN, CNN-LSTM, and an ensemble combining predictions from both a CNN and a CNN-LSTM, with three resampling techniques (RO, SMOTE, and SMOTE + Tomek Links) to classify cardiac arrhythmias. Two open-source datasets, MIT-BIH and PTBDB, were employed, collectively providing around 110,000 heartbeat samples categorized into six distinct classes, with the majority representing normal heartbeats and the remainder denoting various arrhythmia types [119]. The preprocessing stage involved only the computation of the signal derivative, which served as the input to the models. Among the architectures tested, the ensemble model exhibited the best performance, with SMOTE and SMOTE + Tomek Links enhancing the F1-score by 0.31% and 1.63%, respectively, while RO led to a decline in performance [119]. In another study, Sadeghi et al. developed a DNN with a structure consisting of one input layer, three hidden layers, and one output layer to predict diabetes onset [120]. The dataset used was the Tehran Lipid and Glucose dataset, containing demographic, anthropometric, lifestyle, among other features. The DNN was evaluated with four resampling methods: Repeated Edited Nearest Neighbours (RENN), OSS, SMOTE, SVM-SMOTE, and SMOTE-ENN [120]. Experimental findings demonstrated that RENN and OSS achieved the highest F1-scores, while SMOTE resulted in the lowest performance among the tested approaches [120]. In [121], the authors developed a system for identifying human activities in smart homes based on multimodal data, a type of application particularly relevant for assisting elderly individuals or people with cognitive impairments. Specifically, they utilized the SHPERE dataset, which includes data from accelerometers worn by participants, information gathered from Passive Infrared (PIR) sensors, and 2D and 3D data captured by three cameras [121]. The dataset labels corresponded to 20 distinct actions such as climbing stairs, sitting, jumping, and others. After preprocessing and feature extraction, three models—CNN, LSTM, and CNN-LSTM—were evaluated, along with their combinations with resampling techniques, including ROS, SMOTE, and RUS. The results indicated that oversampling methods significantly enhanced accuracy, most notably from 68.29% to 93.67% when SMOTE was combined with the CNN-LSTM model. In contrast, the use of undersampling methods resulted in performance reductions exceeding 50% compared to the baseline with no resampling [121].

The study in [122] introduced a novel framework, DeepSMOTE, which combines an encoder-decoder architecture with the SMOTE algorithm to address class imbalance in image datasets. Raw images are mapped into a low-dimensional latent space via the encoder, where SMOTE is applied to perform oversampling. The generated synthetic samples are subsequently decoded back into the original image space. This approach enhances SMOTE’s ability to process raw images while maintaining high visual quality in synthetic data. Additionally, DeepSMOTE overcomes common limitations of GANs, such as their requirement for extensive data, difficulty in hyperparameter tuning, and vulnerability to mode collapse. The training process for DeepSMOTE begins by encoding and decoding data from all classes to compute a reconstruction loss [122]. To introduce diversity in the latent space, an additional penalty term is incorporated, prompting the network to decode samples that differ from the input images, thereby emulating the SMOTE mechanism in the latent space. This penalty is realized by encoding, shuffling, and decoding samples within the same class, with the reconstruction loss being calculated between the original images and their decoded counterparts derived from different (shuffled) input images. Five publicly available image datasets—MNIST, CIFAR-10, Fashion-MNIST, SVHN, and CelebA—were used for evaluation, alongside six resampling techniques: four pixel-based methods (SMOTE, AMDO, MC-CCR, and MC-RBO) and two GAN-based techniques (BAGAN and GAMO). The ResNet-18 model was chosen as the backbone for all experiments. Analysis of synthetic sample placement showed that DeepSMOTE prioritized generating samples at class boundaries, thereby improving inter-class discrimination. Conversely, GAN-based methods like BAGAN and GAMO concentrated synthetic samples near class centres. In terms of performance, DeepSMOTE significantly outperformed all other methods, particularly when compared to pixel-based oversampling. The study also highlighted DeepSMOTE’s resistance to mode collapse, its robustness to extreme class imbalance, its ability to generate visually realistic synthetic images, and its consistent performance across varied conditions [122]. In a similar approach, Abayomi-alli et al. [113] proposed a method for transforming data into a low-dimensional space, where oversampling is performed using SMOTE-Cov [123]. Subsequently, the data are reverted to their original space for model training. In this case, the dimensionality reduction process is guided by the Kullback–Leibler divergence, aiming to achieve minimal divergence between the original and lower-dimensional representations. Subsequently, reverting the data to the original space is achieved using local regression techniques, such as cubic polynomials, which approximate the mapping based on neighbouring points within the manifold [113]. The proposed algorithm was applied to differentiate between melanoma and non-melanoma skin lesions, yielding an increase in the F1-score from 62.5% to 65.4% on the PH² dataset via resampling, leveraging the SqueezeNet model [113].

Overall, it is apparent that the majority of Deep Learning applications apply resampling directly to raw data, often disregarding complexities such as temporal dependencies in sequential data or spatial patterns in images. Despite this, the examined studies commonly report performance gains, validating the effectiveness of resampling techniques. Nonetheless, it is reasonable to infer that conducting oversampling within lower-dimensional latent space, such as those used in DeepSMOTE, can better preserve data properties, potentially achieving even more pronounced performance improvements. Furthermore, the observed preference for oversampling and hybrid sampling methods, rather than undersampling, is logical given the strong dependency of Deep Learning models on large datasets for effective training. Finally, it is noteworthy that the resampling methods employed are often relatively simple and do not incorporate complexity assessments to adopt more sophisticated resampling protocols, indicates significant potential for optimization in this domain.

After evaluating the three principal categories of resampling methods applied in traditional machine learning and deep learning systems, their main domains of application can be identified. It is important to highlight that most of these techniques are intentionally designed to be application-agnostic, enabling them to handle any datasets with numerical attributes that can either be converted into one-dimensional arrays or are already in the required format. The sole exceptions are DeepSMOTE [122] and the methodology proposed in [113], both of which are explicitly developed for image-related tasks. The reviewed literature reveals the widespread adoption of resampling techniques across several fields. In healthcare applications, examples include Outlier SMOTE for COVID-19 diagnosis [34], Homogeneous Consensus Clustering-Based Undersampling for protein homology prediction [22], emotion detection [118], and cardiac pathology classification [119]. In software engineering, resampling has been employed for bug detection (e.g., SMOTEFUNA [23], CR-SMOTE [63]) and network detection systems [114‐116]. In finance, methods such as Gaussian Mixture Undersampling and SMOTE-ENN integrated with ensemble models leveraging LSTMs and GRUs have been used for credit card fraud detection [117]. These findings are further corroborated by a recent systematic review conducted by Vargas et al., which analysed reported resampling techniques across seven databases [8]. Utilizing the keywords “imbalanced data,” “machine learning,” and “preprocessing,” the researchers reduced a selection of nearly 10,000 articles to 35 after a thorough screening process. Among the selected studies, 34.3% focused on healthcare, 22.9% on finance, 14.3% on engineering, 8.6% on biology, 8.6% on software, and 11.4% on other domains. This distribution reflects and reinforces earlier observations, confirming the cross-disciplinary relevance of resampling methodologies [8].

The comprehensive literature review conducted on this subject matter also demonstrated that the newly proposed methods rarely achieve optimal results across all datasets and under various experimental setups (different datasets with various imbalance ratios, dimensionality, etc.). This holds true even when compared to non-state-of-the-art techniques such as the original SMOTE. To illustrate this observation, Radius-SMOTE exhibited the highest performance in 5 out of 13 datasets [44], GSMOTE-NFM achieved the highest G-mean in 14, 10, and 13 out of 24 datasets when employing SVM [2], Logistic Regression, and C4.5 decision tree classifiers, respectively, while OS-CCD attained the best AUC value in 10 out of the 12 datasets analysed [41].

These findings underscore the notion that, despite the increasing adaptability of resampling approaches to address classification problems, the existence of a singular approach capable of consistently ensuring optimal performance across all scenarios remains uncertain, as would be theoretically expected under the No Free Lunch Theorem [83]. A prominent example of this observation is the efficacy of majority-class based methodologies, such as SWIM, in scenarios characterized by extreme class imbalance, where the available information pertaining to the minority class is deemed insufficient for accurate estimation of its distribution and characteristics [40]. Conversely, when dealing with datasets exhibiting a lower degree of imbalance, minority-based approaches appear to yield superior performance [23]. This discrepancy suggests that, while a given method may demonstrate versatility in handling diverse domains, specific data characteristics, such as extreme imbalance, may benefit from a complete paradigm shift in the fundamental approach employed by the methodologies. Considering these reasoning, a potential alternative for addressing this issue could involve the application of recommendation systems based on complexity metrics or other metrics which provide insights into the resampling methodology with the highest probability of achieving superior performance. Hence several works within this topic are discussed below.

Recommendation systems for resampling methodologies described in scientific literature can be broadly classified into two categories. The first category pertains to studies aiming to establish interpretable relationships or rules between the data characteristics and the resampling strategies, typically derived using data mining techniques such as Exceptional Preferences Mining. Contrarily, the remaining methodologies adopt a more conventional Meta Learning framework, defined as the process of adapting a learning system to a given type of data by utilizing knowledge derived from previous learning tasks involving similar data or data from other domains [125]. Formally, and as described by Parmezan et al. [125], this concept can be defined as follows: Let \(D\) be the problem repository, \(A\) the learning algorithm space, and \(Y\) the performance measure space. Given these assumptions, the goal of a meta-learning system is to map \(f({d}_{i})\) to\(A\), where \(f({d}_{i})\) correspond to a set of meta-features utilized to describe a classification problem (dataset). By doing so, the system can automatically recommend the algorithm \(a\in A\) which maximizes performance \(.\) In practical terms, this type of system must be trained before being integrated into machine learning pipelines. Within this framework, a set of meta-features is typically extracted from each dataset. These meta-features may include statistical properties, such as the number of samples in the dataset, or more complex metrics that provide insights into the presence of data difficulty factors, such as complexity measures. Subsequently, for each dataset, the most suitable resampling method (and potentially its parameterization) is typically determined experimentally, serving as the target variable, enabling meta-model training. Once the meta-classifier has been trained, its incorporation into ML pipelines requires the extraction of the selected meta-features relevant to the specific application. These meta-features are then used to infer the optimal resampling method, which is subsequently applied. A schematic representation of such systems is presented in Fig. 4.

Regarding the first aforementioned approach, in [126], authors utilized 40 datasets sourced from KEEL, UCI, Scikit-learn library, and a medical dataset concerning infections contracted by patients in the Intensive Care Unit, to derive association rules that correlate data characteristics with the corresponding optimal resampling approach. The data characteristics (meta-features) employed included the number of features, number of samples, imbalance ratio, class overlap assessed based on the F1 complexity, and the presence of borderline instances evaluated by the N1 metric. To accomplish this objective, the authors initially evaluated the resampling technique (RUS, ROS [24], SMOTE [23], SMOTE + OSS, SMOTE + CNN, SMOTE + ENN [58], and SMOTE + TL [45]) that ensured the highest performance for each dataset, as measured by the Index of Balanced Accuracy (IBA). This measure was obtained through a tenfold cross-validation process. Following the creation of the meta-dataset, the Apriori Algorithm was employed for mining association rules, enabling the derivation of rules such as: if the imbalance ratio is less than 48, N1 ranges from 6.75 to 12.56, and F1 exceeds 78.8, then ROS is the most suitable resampling approach. The analysis of the obtained rules was consistent with prior results reported in the literature, indicating that no resampling produced superior results if the imbalance ratio was extremely low. Similarly, it was observed that SMOTE, when combined with CNN, OSS, and TL, was the most frequently chosen approach according to the derived rules, highlighting the effectiveness of hybrid methods. Moreover, other rules were induced not based on the IBA value, but from the rankings of resampled approaches, as assessed by Friedman and post hoc statistical tests. Lastly, it is worth noting that, despite the study's ability to infer rules about the relationship between data characteristics and suitable resampling approaches, it did not assess the model's generalizability on new data, thereby limiting the evaluation of the applicability of the drawn conclusions.

Alternatively, Luengo et al. attempted to evaluate the potential of complexity metrics in setting the optimal operating ranges for three resampling methods, namely SMOTE [23], SMOTE-ENN [58], and EUSCHC [102], by utilizing 44 datasets from the UCI repository [127]. The overarching aim was to establish a system that, given the characterization of classification problems using complexity measures, could predict the suitability of a specific resampling method. In this regard, a resampling method was considered to have a good behaviour if the test AUC, evaluated through fivefold cross-validation, surpassed 0.8, and the difference between the training and test AUC was less than 0.1. In contrast, a method was classified as having a bad behaviour if it did not meet one of these conditions. In practice, the determination of optimal intervals entailed a graphical representation of the AUCs yielded for each dataset, organized in an ascending manner according to the corresponding complexity value. Such an approach allowed for the identification of areas populated by consecutive values of bad or good behaviour, enabling the determination of the competence domains for each method. Through this procedure, the complexity measures F1, N4, and L3 proved capable of extracting the most significant intervals [127].

Lastly, we highlight the methodology suggested by Santos et al., where Exceptional Preferences Mining is used to procure interpretable rules implying the suitability of specific methods over others [128]. From a methodological perspective, this approach entailed the construction of a meta-dataset, where meta-features were extracted via the Python library pymfe [129] (namely statistical, complexity, landmark, among other features), and the target variables corresponded to the rankings of 9 resampling approaches based on the F1-score. The utilized resampling techniques were the following: ROS, SMOTE [102], SafeLevel-SMOTE [31], Borderline-SMOTE [28], ADASYN [30], AHC [103], ADOMS [130], SMOTE-TL [45], and SMOTE-ENN [58]. The evaluation of the results endorsed some claims made in "Class imbalance: formal definition" as well as the previously discussed work of Kraiem et al. [126], as it was observed that if a dataset's complexity is minimal, reflected by lower values of L2, N1, N4, and T3, or a higher percentage of safe points versus borderline points, the implementation of resampling approaches becomes redundant [128].

Addressing the second typology of methods previously mentioned, a study by Morais et al. advocated the utilization of simple attributes (such as the number of samples, attributes, presence of outliers, etc.) and statistical features to ascertain the most suitable undersampling method and its optimal parameterization, employing a k-NN based approach [131]. Practically, the recommendation process involves calculating the meta-features for the dataset under analysis, followed by the detection of the nearest neighbour amongst the existing meta-examples. The system’s output is the ranking of algorithms associated with the nearest neighbour. The undersampling techniques incorporated into this recommendation system were ENN [77], KMUS, MD, NCL [77], NearMiss-1 [78], NearMiss-3 [78], and RUS. However, varying parameterizations were tested for each resampling method, culminating in a total of 491 candidate resampling algorithms that could be chosen for an imbalanced input dataset. Experimentally, the authors utilized 29 publicly accessible datasets, and the optimal methods for each dataset were identified using a weighted performance metric that included Accuracy, AUC, F1-score, Specificity, and Negative Predicted Value. By evaluating the results, it was evident that the proposed system consistently outperformed a Random search approach, which involves choosing the best method from 50 randomly selected resampling methods and their corresponding parameterizations. Furthermore, it registered equal performance of a brute force search in 8 out of the 29 datasets [131]. In [132], the authors adopted a very similar approach based on k-NN rank aggregation to generate recommendations. However, the developed system was not only able to recommend resampling approaches, but also algorithmic level approaches, such as Bagging, AdaBoost [133], MetaCost [134], SMOTEBoost [135], among others. In addition, other meta-features were used, namely the complexity metrics proposed by Ho and Basu [136], Landmarking measures, model-based measures and structural measures. A thorough testing of the developed system on 80 unbalanced binary datasets from UCI, KEEL and Promise repositories yielded excellent results, specifically that in 96% of the times, the recommended approaches are within the top three techniques.

As we conclude this section, it is important to note that these systems are built upon a limited number of datasets, generally smaller in size and frequently corresponding to binary classification problems, thus restricting their capacity for generalization and integration into machine learning pipelines of various domains. In addition, in light of the discussion concerning the limitations of the Imbalance Ratio in predicting classification task difficulty and the bias of complexity measures in imbalanced regimes, it is likely that these systems could be enhanced with metrics that can more precisely assess the complexity of the classification task, such as those presented in "Classification complexity in imbalanced domains". Despite all the aforementioned aspects, this approach holds considerable potential due to its reduced computational cost and the ability to avoid a brute-force search when determining the optimal methodology, which seems to be the most common approach [131]. Furthermore, these systems obviate the need for users to possess deep technical knowledge about the most fitting resampling techniques, thus facilitating the application of these approaches in other areas.

Lastly, it is worth noting that, outside recommendation systems, there may be other viable approaches to tackle this issue such as the utilization of ensemble resampling methods to mitigate performance variance in different datasets. In [22], the authors proposed the application of cluster-based undersampling using different algorithms and parameterizations to minimize the influence of selected parameters, which have been demonstrated to impact performance. Moreover, the integration of complexity-based recommendation systems could prove valuable in determining the relative importance of base resampling methods based on the specific characteristics of the data.