Penalized partial least square discriminant analysis with ℓ1-normℓ1-norm for multi-label data

• A new and effective learning method for the multi-label data is proposed.
• The method exploits the sparse property of the label space fully.
• The method captures the correlations of the variables by using partial least squares discriminant analysis.
• Dimension reduction has also applied to the high-dimensional data.
• The sparse purpose is achieved by performing the l1-norm penalty.

Multi-label data are prevalent in real world. Due to its great potential applications, multi-label learning has now been receiving more and more attention from many fields. However, how to effectively exploit the correlations of variables and labels, and tackle the high-dimensional problems of data are two major challenging issues for multi-label learning. In this paper we make an attempt to cope with these two problems by proposing an effective multi-label learning algorithm. Specifically, we make use of the technique of partial least square discriminant analysis to identify a common latent space between the variable space and the label space of multi-label data. Moreover, considering the label space of the multi-label data is sparse, a l1-norm penalty is further performed to constrain the Y-loadings of the optimization problem of partial least squares, making them sparse. The merit of our method is that it can capture the correlations and perform dimension reduction at the same time. The experimental results conducted on eleven public data sets show that our method is promising and superior to the state-of-the-art multi-label classifiers in most cases.