Unsupervised feature selection based on local structure learning

Yanbei Liu, Lei Geng, Fang Zhang, Jun Wu, Liang Zhang, Zhitao Xiao

Abstract: Unsupervised feature selection has become a significant and ambitious issue due to vast amounts of high-dimensional unlabeled data in machine learning. Traditional unsupervised feature selection algorithms usually make use of the similarity matrix for feature selection, and they heavily rely on the learned structure. However, a large amount of actual data always contains many noise samples or features that may make the similarity matrix obtained from the original data unreliable. Using Ideal Local Structure Learning (LSL) method ,we propose a novel unsupervised feature selection to perform feature selection and local structure learning at the same time in this paper. In order that we can earn more exactly structure information, an ideal local structure with precisely c connected components of data (c is the number of clusters) is utilized to refine the similarity matrix. Moreover, in order to optimize our algorithm, an effectual and plain iterative algorithm is developed. Experiments on multiple public baseline datasets, including biomedical data, letter recognition digit data and face image data, reveals the outstanding performance of our algorithms in the most advanced aspects.