K-IPO: Preserving Feature Importance in Imbalanced Data Oversampling.

Oversampling is a common technique to address class imbalance in tabular classification datasets. However, existing oversampling methods often inadvertently distort the feature importance ranking, which is crucial for model interpretability and explanations. This distortion can lead to misleading insights about which features are truly driving a model's predictions. To tackle this, researchers have developed Kendall-constrained Importance-Preserving Oversampling (K-IPO). This generator-agnostic framework operates by iteratively generating minority-class candidates and accepting them only if their inclusion maintains a user-defined minimum Kendall's tau correlation with the original data's feature importance ranking. Optional stricter constraints can be applied to the most highly ranked features. Evaluated across 20 imbalanced binary classification datasets with various classifiers and explanation methods, K-IPO consistently demonstrated superior or competitive performance. It excelled in preserving feature importance, maintaining explanation consistency, and improving class separability. Furthermore, K-IPO generally enhanced predictive performance while keeping computational overhead manageable, offering a significant advancement for interpretable AI on imbalanced data.