A Filter-based Unsupervised Feature Selection Method via Improved Local Structure Preserving

In this paper, we propose a novel filter based unsupervised feature selection algorithm. We first extract the global level manifold structure using LLE on all features. We also extract the feature level manifold structure using LLE on each single feature. We then compute the feature-wise non-negative local linear reconstruction weight to capture the feature relationship. The true manifold structure of feature is then computed by the linear combination of its own Laplacian matrix and its neighbor's Laplacian matrices. The importance of feature is then evaluated by the difference between the global manifold structure and the combined feature level manifold structure. Extensive experimental results on benchmark data sets well demonstrate that the proposed method outperform state-of-the-art filter-based unsupervised feature selection methods.