Locally linear embedding vote: A novel filter method for feature selection

Abstract

Feature selection algorithm based on the locally linear embedding (LLE) cannot effectively detect the changes in the neighborhood of data and cannot take full advantage of the graph-preserving ability. To address these issues, a novel filter method integrated LLE, named LLE vote, is proposed. In the proposed algorithm, we first construct two graphs of each feature through computing the local structure and sparse structure of data in high-dimensional manifold, and then a novel metric criterion is employed to rank the features by measuring the difference between the reconstructed feature and the corresponding original input. Extensive experiments are carried out on benchmark fault data set and the other kind of data from our own laboratory, and the experimental results not only demonstrate the effectiveness of the proposed method, but also indicate that the LLE vote outperforms the existing state-of-art methods.