Sparse regularized discriminative canonical correlation analysis for multi-view semi-supervised learning

For multi-view data representation learning, recently the traditional unsupervised CCA method has been converted to supervised ways by introducing label information from samples. However, such supervised CCA variants require large numbers of labeled samples which hampers its practical application. In this paper, in order to mine the most discriminant information only from a few labeled samples, inspired by sparse representation we propose a novel sparse regularized discriminative CCA method to make use of the label information as much as possible. Through constructing sparse weighted matrices in multiple views, we incorporate the structure information into the original CCA framework to extract fused multi-view features which not only are the most correlated but also carry the important discriminative structure information. Our approach is evaluated on both handwritten dataset and face dataset. The experimental results and the comparisons with other related algorithms demonstrate its effectiveness and superiority.