Semi-supervised dimensionality reduction based on global and local scatter

Recently, semi-supervised learning, which uses unlabeled samples and supervised information together in learning process, has received much attention. Compared with class labels, pairwise constraints is a kind of supervised information which are more easily to obtain. In this paper, a new algorithm is proposed, called as SSGL, which preserves both the global (intrinsic) and local structure of the original data (both labeled and unlabeled data) and incorporates the structure of the pairwise constraints in the projected low-dimensional space. In this way, the novel algorithm intends to find a better discriminative projection. Experiment results on Yale and ORL face databases show its effectiveness.