Multi-label Learning by Exploiting Imbalanced Label Correlations

Abstract

Multi-label classification refers to the supervised learning problem where an instance may be associated with multiple labels. It is well known that exploiting label correlations is important for multi-label learning. Existing approaches typically assume that the distribution of classes is balanced. In many real-world applications, multi-label datasets with imbalanced class distributions occur frequently, which may make various multi-label learning methods ineffective. Since the existing multi-label learning algorithms pay less attention to the problem of correlation with imbalanced label sets, this paper proposed a Multi-Label learning model by exploiting Imbalanced Label Correlations (ML-ILC). ML-ILC uses graph convolution neural network to learn the correlation between labels. At the same time, we suggest that the regularization of minority classes is stronger than that of frequent classes, which can improve the generalization error of minority classes. To investigate the performance of the proposed multi-label learning model, we considered two benchmark datasets including VOC2007 and COCO. The proposed method successfully achieved better classification performance compared to the state-of-the-art compression methods.