Multi-label Learning with Data Self-augmentation

Abstract

Multi-label learning (MLL) deals with the problem where each training example is associated with multiple labels. Existing MLL approaches focus on manipulating feature space and modeling label dependencies among labels. However, both of them require additional burden assumps and cannot easily be embedded into existing algorithms, while data augmentation is the more intuitive way to facilitate MLL. Therefore, in this paper, we propose a novel data augmentation method named MLAUG, i.e. Multi-Label learning with data self-Augmentation for MLL. Specifically, to achieve data augmentation, we learn feature correlation and label correlation matrices in feature space and label space simultaneously in an adaptive manner. Guided by the learned correlation matrices, MLAUG is able to refine the original feature and label spaces by linear combinations of other data vectors. In this way, we could obtain semantic-richer feature distribution and smoother label distributions, thus facilitating multi-label predictive performance. Besides, to further improve the model’s performance, we introduce feature and label graph Laplacian regularization for guaranteeing the discriminability and capturing more adequate label correlation, respectively. Extensive experiment results demonstrate the effectiveness of our proposed MLAUG.