Semi-supervised Subspace Learning via Constrained Matrix Factorization

This paper adopts the matrix factorization approach by improving the NMF model to build a semisupervised learning framework (DCNMF) that integrates the linear discriminate analysis ( LDA) and base cone volume constraints. The proposed DCNMF is a subspace learning model in which minimizing the basic cone volume of the learned subspace and reducing the data dimensionality so that the distance within-class samples is minimized and the distance between-class samples is maximized. The proposed method is evaluated by experiments on two cases of face recognition tasks, namely, various numbers of training data and different subspace dimensionalities.