Adaptive sample repulsion against class-specific counterfactuals for explainable imbalanced classification.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Imbalanced classification presents challenges in complex feature spaces with overlapping sample regions.
• Existing methods often fail to deeply model feature-label relationships or provide instance-level explanations.
• This limits improvements in classification performance and model credibility.

Purpose of the Study:

• To propose an explainable imbalanced classification framework (CSCF-SR) that dynamically regulates feature-space distribution.
• To form a closed-loop between explanation generation and classification decisions using counterfactual samples.
• To enhance model classification capability for samples in overlapping regions.

Main Methods:

• A class-specific dual-actor reinforcement learning architecture for counterfactual searching.
• A multi-step dynamic perturbation mechanism for precise counterfactual sample generation.
• Adaptive sample repulsion exploiting displacement vectors to clarify class boundaries.

Main Results:

• CSCF-SR demonstrated superior performance over 27 imbalanced classification methods on F1-score and G-mean across 50 datasets.
• Significant improvements were observed on 25 datasets with severe class overlap.
• The framework effectively enhances classification for samples within overlapping regions.

Conclusions:

• The proposed CSCF-SR framework offers a novel approach to imbalanced classification by integrating explainability and adaptive sample manipulation.
• The method shows significant performance gains, particularly in challenging scenarios with high class overlap.
• This work contributes to more credible and accurate imbalanced classification models.