Oversampling the Minority Class in the Feature Space.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Real-world datasets often exhibit class imbalance, posing significant challenges for machine learning algorithms.
• Traditional oversampling methods may not effectively capture the underlying data distribution.

Purpose of the Study:

• To propose and evaluate a novel synthetic oversampling technique operating in kernel-induced feature spaces.
• To enhance the performance of machine learning models on imbalanced datasets.

Main Methods:

• Synthetic oversampling in the empirical feature space (EFS), which is isomorphic to the kernel-induced feature space.
• Investigation across three scenarios: full and reduced-rank EFS, kernel learning for class separation, and a unified preferential oversampling framework.
• Extensive experimentation on 50 imbalanced datasets, particularly within the context of Support Vector Machines (SVMs).

Main Results:

• Demonstration of synthetic oversampling in feature space as an effective strategy for imbalanced data.
• Analysis of the influence of kernel function choice on the structure of the feature space and oversampling effectiveness.
• Validation of the proposed unified framework's ability to encompass existing oversampling approaches.

Conclusions:

• Kernel-induced feature space oversampling offers a promising direction for tackling class imbalance.
• The method provides a flexible and powerful tool for improving machine learning model generalization on imbalanced data.
• The empirical feature space (EFS) serves as a practical environment for implementing these advanced oversampling techniques.