Geometrically Preserved Dual Projections Learning for Multi-label Classification

Multi-label classification encounters great challenges in processing high dimensional input feature and output label spaces. In this paper, we propose a geometrically preserved dual projections learning of both label and feature spaces for multi-label classification. Different from traditional multi-label classification algorithms which learn appropriate feature dimensions directly from feature space, our proposed method learns dual label and feature space projections from both input feature and output label spaces simultaneously. Thus, achieving dimension reduction in both label and feature spaces. Furthermore, since labels and features are sampled from label and feature manifolds, respectively, we learn their geometric structures by constructing label and feature graphs. By geometrically preserving dual projections, better low rank structures in label and feature spaces are obtained, which further improves the performance of multi-label classification. Experimental results on several data sets such as cal500, enron, languagelog and rcv1s1, demonstrate the proposed method outperforms the state-of-the-art multi-label classification methods, such as MIFS, RAKEL, PLST, CPLST, GroPLE, CSSP, FaIE in LSDR and STOA.