Weight Pairwise Constraints Based Semi-supervised Locality Dimensionality Reduction Algorithm

Abstract

The existing semi-supervised dimensionality reduction algorithms regard the side information as equal, which can not fully utilize the side information. A novel algorithm called weight pair wise constraints based semi-supervised locality dimensionality reduction is proposed. The algorithm expands the side information by k-nearest neighborhood graph, which can increase the number of the constraints. The side information is integrated into the neighborhood graph to revise it and the constraints are propagated to the neighbors at the same time. Then each pair wise constraint is weighted by its information power. A projection is found based on the revised neighborhood graph and the weighted constraints, the projection not only maintain the local geometric structure but also decrease the distance within-class and increase the distance between-class. The experimental results on UCI datasets show the effectiveness of the algorithm.