Hybrid density-based adaptive weighted collaborative representation for imbalanced learning

Yanting Li, Shuai Wang, Junwei Jin, Hongwei Tao, Chuang Han, C. L. Philip Chen

Abstract: Collaborative representation-based classification (CRC) has been extensively applied to various recognition fields due to its effectiveness and efficiency. Nevertheless, it is generally suboptimal for imbalanced learning. Previous studies have revealed that a class-imbalance distribution can lead CRC, and even most conventional classification methods, to ignore the minority class and prioritize the majority class. To address this deficiency, this paper proposes a hybrid density-based adaptive weighted collaborative representation model that incorporates a regularization technique and an adaptive weight generation mechanism into the CRC framework. A new regularization term, based on class-specific representation, is introduced to decrease the correlation between classes and enhance CRC’s discriminative ability. The sample distribution and density information within and between classes are employed to assign greater weights to minority samples, thereby strengthening the representation capabilities of minority samples and reducing the bias towards the majority class. Furthermore, it is theoretically demonstrated that this model has a closed-form solution. Its complexity is comparable to that of CRC, ensuring its efficiency. Extensive experiments on diverse data sets from the KEEL repository show the superiority of the proposed method compared to other state-of-the-art imbalanced classification methods.