Local Density-Aware Oversampling with Noise Resistance for Imbalanced Data Classification

Imbalanced data classification has become a critical challenge in the field of machine learning. Traditional oversampling approaches often suffer from synthetic sample inaccuracy or aggravated class overlap problems. To better address these challenges, this paper proposes an oversampling method that integrates noise detection with adaptive sample generation. Specifically, we first introduce a noise detection strategy based on the average $k$-nearest neighbor distance, which identifies and removes high-interference noisy samples through local density analysis. Next, we design a weight allocation mechanism that jointly evaluates each instance's boundary risk and generation potential, prioritizing the synthesis of higher-weighted samples. Finally, to better preserve the classification boundary, we incorporate a neighbor class-sensitive coefficient into the sample generation process. Extensive experiments on 17 benchmark datasets demonstrate that the proposed method significantly outperforms well-known oversampling-based approaches, achieving superior classification performance.