Dimension Reduction Using Nonnegative Matrix Tri-Factorization in Multi-label Classification

Multi-label classification problem has become more important in image processing and text analysis where an object often is associated with many labels at the same time. Recently, even in this problem setting dimension reduction aiming at avoiding the curse of dimensionality has gathered an attention, but it is still a challenging problem. Nonnegative Matrix Factorization (NMF) is one of promising ways for dimension reduction in unsupervised learning, and is extended from two-matrix factorization to triple-matrix factorization. In this paper, we reformulate the NMF with three factor matrices in such a way that it is solvable the problem of the combinatorial explosion of labels and incorporates the label correlation naturally in supervised learning. Experiments on web page classification datasets show the advantages of the proposed algorithm in the classification accuracy and computational time.