A Combined Algorithm for Imbalanced Classification Based on Dual Distribution Representation Learning and Classifier Decoupling Learning

Abstract

Existing classification algorithms for imbalanced datasets adopt data resampling, classes reweighting and other class balancing strategies to strengthen representation ability for minority classes and adjust the classification interface. However, these algorithms weaken the network’s representation ability for majority classes. Therefore, a combined algorithm is proposed based on dual distribution representation learning (DDRL) and classifier decoupling learning (CDL). Here, DDRL preserves the original distribution and samples the balanced distribution from it to guide the learning of dual distribution representation, which enhances minority classes' feature representation ability and retains it for majority classes. CDL decouples the classifier from feature representation network, and trains an MLP classifier with a balanced subset, aiming at adjusting the classification deviation caused by weak features of minority classes. Experimental results show that the proposed algorithm can improve the classification accuracy on class imbalanced datasets effectively.