An efficient multi-label learning method with label projection

Abstract Multi-label classification (MLC) is a problem that each given sample is associated with more than one label simultaneously. There is a variety of application in our daily life, such as text categorization and image annotation. To date, many methodologies are proposed to do a multi-label learning task. According to the MLC setting, we put forward an MLC method called TPMLC (an MLC method with Two Parts) and propose a uniform loss function based on the variational inference with Bayesian and Gaussian distribution assumption. Moreover, this uniform loss function is composed of two parts. On one hand, the first part is about the determination of the relationship between sample and multiple labels, so we adopt a set of multiple support vector machines (SVMs) to determine this relationship. On the other hand, the second part in this uniform loss function is about the determination of the relationship among multiple labels, and thus we construct a projection matrix model to determine this relationship. Furthermore, this uniform loss function is optimized simultaneously, such that the two kinds of relationships can be optimized at the same time. Besides, we also present the convergence analysis and computational complexity analysis of the method. After that, in the experiment part, the comparison of TPMLC with state-of-the-art approaches manifests the feasibility and the competitive performance in classification. In addition, the statistic results show that the proposed method performs better than the state-of-the-art methods.