Iterative Laplacian Score for Feature Selection

Laplacian Score (LS) is a popular feature ranking based feature selection method both supervised and unsupervised. In this paper, we propose an improved LS method called Iterative Laplacian Score (IterativeLS), based on iteratively updating the nearest neighborhood graph for evaluating the importance of a feature by its locality preserving ability. Compared with LS, the key idea of IterativeLS is to gradually improve the nearest neighbor graph by discarding the least relevant features at each iteration. Experimental results on several high dimensional data sets demonstrate the effectiveness of our proposed method.