Robust Locality Preserving Projections with Pairwise Constraints

Dimensionality reduction is one of the key processes of high dimensional data analysis, including machine learning and pattern recognition. Constrained Locality Preserving Projections (CLPP) is a variant of Locality Preserving Projections (LPP) plus with pairwise constraints and constraints propagation. Like LPP, however, CLPP is still sensitive to noise and parameters. To overcome these problems, we propose a Robust LPP (RLPP) with pairwise constraints based on robust path based similarity. In contrast to CLPP, RLPP is less sensitive to noise and outliers. Besides, it capitalizes on pairwise constraints more efficiently. Experimental results on real world datasets confirm its effectiveness.