Multi-view label space dimension reduction

Abstract

In multi-label classification, the explosion of the label space makes the classic multi-label classification models computationally inefficient and degrades the classification performance. To alleviate the curse of dimensionality in label space, many label space dimension reduction (LSDR) algorithms have been developed in last few years. Whereas, they are all designed for single-view learning and ignore that one sample can be represented from different views. In this paper, we propose a multi-view LSDR model for multi-label classification. The weights of different views are learned and then multi-view label embedding results are combined by the learned weights. Experiments on benchmark datasets show that the proposed multi-view learning model outperforms the best single-view model and the majority voting method.