A robust feature selection criterion based on preserving locality structure

Feature selection plays an important role in seeking of better representation for pattern recognition tasks. Stability of selected feature is a measurement of generalization ability of feature selection algorithm. In this paper, we propose a novel feature selection criterion based on Preserving the Locality Structure (PLS) of image database in feature spaces. The basic idea is to introduce the locality relationship of classes into the feature selection criterion. To validate the proposed PLS criterion, we take face recognition as the application example and perform experimental comparison with several classic feature selection strategies such as the Fisher Score (FS) criterion and the minimal redundancy maximal relevance (mRMR) criterion. Experimental results show that the proposed PLS criterion could result in close or even better accuracy as the mRMR with much less computation cost. Moreover, the selected features with PLS are very stable compared with the classic criterions. © 2013 Binary Information Press.