Supervised low dimensional embedding for multi-label classification

Abstract

In multi-label classification, discovering label structures or label correlations when learning can improve predictive performance and time complexity. In this paper, a unified framework is proposed to incorporate the supervised correlation exploration with the predictive model. In the framework, feature mappings to a low-dimensional subspace is obtained through a linear transformation guided by the label information. And a multi-label classifier is simultaneously built on the projected features. The framework leads to a trace optimization problem which can be solved by a generalized eigenvalue problem. Meanwhile, the dual form of the framework is presented to deal with different problems. Experiments on four datasets show that the proposed framework can achieve comparable performance with four other well-known methods, and achieve better performance when label correlations are important. It's also demonstrated that the framework is efficient when the dimensionality is low, and the dual form will be more efficient without extra computational tricks in the small-sample problems.