AxelSMOTE: An Agent-Based Oversampling Algorithm for Imbalanced Classification
This work addresses the problem of skewed datasets hindering minority class performance for machine learning practitioners, presenting a novel method that is incremental in improving upon existing oversampling techniques.
The paper tackles class imbalance in machine learning by introducing AxelSMOTE, an agent-based oversampling algorithm that addresses drawbacks of traditional methods, and it outperforms state-of-the-art sampling methods on eight imbalanced datasets while maintaining computational efficiency.
Class imbalance in machine learning poses a significant challenge, as skewed datasets often hinder performance on minority classes. Traditional oversampling techniques, which are commonly used to alleviate class imbalance, have several drawbacks: they treat features independently, lack similarity-based controls, limit sample diversity, and fail to manage synthetic variety effectively. To overcome these issues, we introduce AxelSMOTE, an innovative agent-based approach that views data instances as autonomous agents engaging in complex interactions. Based on Axelrod's cultural dissemination model, AxelSMOTE implements four key innovations: (1) trait-based feature grouping to preserve correlations; (2) a similarity-based probabilistic exchange mechanism for meaningful interactions; (3) Beta distribution blending for realistic interpolation; and (4) controlled diversity injection to avoid overfitting. Experiments on eight imbalanced datasets demonstrate that AxelSMOTE outperforms state-of-the-art sampling methods while maintaining computational efficiency.