An ensemble contrastive classification framework for imbalanced learning with sample-neighbors pair construction

While existing imbalanced classification methods have made great progress, there are still some challenges in the current imbalanced learning field: (1) How to achieve the balance between classes without introducing noise and losing information; (2) How to mine the differences between classes from datasets with a relatively small number of positive samples; (3) How to learn the distribution differences in overlapping areas efficiently. To address these problems, this paper proposes an ensemble contrastive classification framework with sample-neighbors pair construction. The traditional pointwise imbalanced classification is redefined as a pairwise label matching task between the sample to be classified (Target Sample, TS) and a group of neighbor samples (Contrastive Sample Group, CSG). For any TS, we can obtain multiple CSGs containing same/different class with the TS, and combine the TS with different CSGs into sample-neighbors pairs as positive/negative samples in the new task. In this way, the balance of classes and multiplied increase of sample scale are achieved without introducing any noise. Based on the obtained rich data, a robust classifier can be trained to mine the distribution differences in overlapping areas through the contrastive learning between TS and its CSGs. For a given test sample, we can obtain abundant sample-neighbors pairs and their corresponding classification results. Its label can be obtained by result integration and reverse reasoning. Extensive experimental evaluations on 48 KEEL and UCI public datasets show that the proposed method outperforms the existing state-of-the-art imbalanced classification methods in terms of F1-measure and G-mean.