Discriminative locality preserving dimensionality reduction based on must-link constraints

Locality preserving manifold learning algorithms are seeking intrinsic manifold based on overlapping local geometry structure. Locality Preserving Projections (LPP) and Neighborhood Preserving Embedding (NPE) are two representative linear locality preserving manifold learning algorithms, which not only defined on training samples, but also can generalize to test samples. But they just take the local structure into consideration, ignoring some available prior information. Pairwise constraints are easier obtained supervised information compared with labels. In this paper, we proposed two discriminative locality preserving manifold learning algorithms, by incorporating must-link constraints into LPP and NPE to improve their discriminative ability. Experiments results on Yale and ORL face databases verified the effectiveness.