A cluster-assisted differential evolution-based hybrid oversampling method for imbalanced datasets.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Class imbalance in datasets poses a significant challenge, leading to biased machine learning models that misclassify minority class instances.
• Existing oversampling techniques like SMOTE often struggle with issues such as class overlap, poor decision boundary representation, and noise accumulation.

Purpose of the Study:

• To introduce ClusterDEBO, a novel hybrid oversampling method integrating K-Means clustering and differential evolution (DE).
• To generate synthetic samples in a structured and adaptive manner, improving the handling of imbalanced datasets.

Main Methods:

• The method partitions minority class data into clusters using the silhouette score to determine the optimal number of clusters.
• Differential evolution-based mutation and crossover operations generate diverse synthetic samples within each cluster, preserving data distribution.
• A selective sampling and noise reduction mechanism filters synthetic samples based on their classification performance impact.

Main Results:

• ClusterDEBO was evaluated on 44 benchmark datasets using kNN, DT, and SVM classifiers.
• The proposed method consistently outperformed existing oversampling techniques, enhancing class separability and classifier robustness.
• Statistical validation using the Friedman test confirmed the significance of the observed improvements.

Conclusions:

• ClusterDEBO offers a powerful strategy for handling imbalanced datasets by leveraging cluster-assisted differential evolution.
• The method demonstrates superior performance in improving classifier accuracy and robustness compared to traditional oversampling techniques.