DENI: A Density-Enhanced Hybrid Sampling Framework with Neighborhood Information for Noisy Imbalanced Classification

The class-imbalance problem is a prevalent challenge in real-world applications. Label noise aggravates the complexity of classification. Conventional classifiers tend to learn a bias toward the majority class, which results in poor performance on the minority class. For noisy imbalanced data, the resampling method is one of the most competitive methods among the existing techniques. However, it is susceptible to improper sampling strategies. To this end, this paper proposes a density-enhanced hybrid sampling framework with neighborhood information (DENI). DENI first introduces the global density gap of the dataset to distinguish the noisy samples, which can reduce the impact of noise on classification. Second, a density-enhanced oversampling process with neighborhood information is designed, which enables the potential regions of the minority class to be appropriately expanded without generating the scattered samples or compressing the minority-class regions. Finally, random rebalance is implemented to obtain a balanced dataset. Extensive experiments on diverse benchmark noisy imbalanced datasets demonstrate the effectiveness and robustness of the proposed DENI.