Stepwise optimal feature selection for data dimensionality reduction

In this paper, an unsupervised feature selection algorithm, stepwise optimal feature selection (SOFS), is introduced for data dimension reduction. The feature subset is obtained through a two-step selecting approach. Firstly the global optimal feature is selected according to the similarity graph constructed by the pair-wise feature similarity matrix, and then the features with furthest distances to subset are chosen step by step. The selection process is terminated when the optimal feature is nearer to the selected subset than to the residual features. Experiments carried on the UCI datasets show that SOFS algorithm gets sound performance in speed and obtains higher classification accuracy rates than state-of-art method using two popular classifiers, and SOFS sharply reduces data dimensionality with slightly weakening the classification power. Copyright © 2015 Binary Information Press.