PPML: Penalized partial least squares discriminant analysis for multi-label learning

Abstract

Multi-label learning has attracted widespread attention in machine learning, and many multi-label learning algorithms have been witnessed. However, two main challenging issues remain: the high dimension of data and the label correlation. In this paper, a new classification method, called penalized partial least squares discriminant analysis for multi-label learning (PPML), is proposed. It aims at performing dimension reduction and capturing the label correlations simultaneously. Specifically, PPML first identifies a latent space for the variable and label space via partial least squares discriminant analysis (PLS-DA). To tackle with the problem of high dimensionality in solving PLS-DA, a ridge penalization is exerted on the optimization problem. After that, the latent space is used to construct learning model. The experimental results on the standard public data sets indicate that PPML has better performance than the state-of-the-art approaches. © 2014 Springer International Publishing Switzerland.