Fast Backward Iterative Laplacian Score for Unsupervised Feature Selection

Iterative Laplacian Score (IterativeLS), an extension of Laplacian score (LS) for unsupervised feature selection, iteratively updates the nearest neighborhood graph for evaluating the importance of a feature by its local preserving ability. However, LS and IterativeLS separately measure the importance of each feature and do not consider the association of features. To remedy it, this paper proposes an enhanced version of IterativeLS, called fast backward iterative Laplacian score (FBILS). The goal of FBILS is to fast remove some unimportant features by taking into account the association of features. The proposed FBILS evaluates the feature importance according to the joint local preserving ability that reflects the association of features. In addition, FBILS deletes more than one feature in an iteration, which would speed up the process of feature selection. Extensive experiments are conducted on UCI and microarray gene datasets. Experimental results confirm that FBILS can achieve a good performance.