Sparse and low-rank representation for multi-label classification

Multi-label learning deals with the problem where each instance may be associated with multiple labels simultaneously, and how to discover and exploit the label correlations is one of important research issues. In this paper, we propose a novel sparse and low-rank representation-based method for multi-label classification (SLMLC), which can automatically exploit the asymmetric correlations among labels while learning the model parameters in a unified learning framework. More specifically, we assume that the weight matrix is divided into a sparse matrix and a low-rank matrix, where the sparse and low-rank matrices are utilized to capture the specific features that are relevant to each label and the shared feature subspace among all labels, respectively. Then, we integrate multi-label classification and label correlations into a joint learning framework to learn the correlations among labels and the model parameters simultaneously. Lastly, the formulation is transformed into its convex surrogate due to its non-convexity, and we solve it by developing an alternating iterative method. Experimental results on fifteen data sets in terms of six evaluation criteria show that SLMLC achieves superior performance compared to the state-of-the-art multi-label classification algorithms.