An imbalanced contrastive classification method via similarity comparison within sample-neighbors with adaptive generation coefficient

Correct discrimination of samples in overlapping regions is crucial in imbalanced classification problems. Data-level methods generate new samples in overlapping areas to obtain a clearer classification boundary. However, the generated samples' reliability cannot be guaranteed and additional noise will be introduced. Recently, although a few researchers have introduced contrastive learning to address the above problems, they have neither explored the differences in information content of samples in the contrastive task, nor considered the complex samples in overlapping areas. This paper proposes a contrastive classification method based on the similarity comparison of sample-neighbors, which transforms the traditional label prediction task into a similarity analysis task. Considering the distribution of neighbor category and the information content in the comparison task, each sample's unique generation coefficient is calculated. On this basis, a similarity loss with the target-neighbor sample group is designed so that the model can calculate the similarity between different samples. Meanwhile, extra discriminator will supervise the generated samples of variational autoencoder (VAE), which prompts the model to focus on the characteristics of individual samples. Experimental results on 39 public datasets show that the proposed method outperforms typical imbalanced classification methods.