Robust unsupervised feature selection via data relationship learning

Abstract

Unsupervised feature selection robust to many outliers is a challenging task. The crucial difficulty is learning a robust subspace, which preserves local structure. The most common solution is to reduce fitting error by applying different robust norms. However, there are three shortcomings. Firstly, they are not robust enough when outliers distributed both randomly and concentratedly are widely present. Secondly, outlier removal is not considered. Thirdly, it is not easy to understand and choose an euclidean distance threshold that decides a sample as an outlier in different scenarios. The first two shortcomings make previous methods fail to achieve their expected learning results, and the third one increases the application difficulty in different fields. To address these issues, a robust unsupervised feature selection via data relationship learning (RUFSDR) is proposed in this paper. Specifically, scores representing the data’s importance will be learned and assigned to each sample. Inliers will be given different positive scores. Outliers will be given 0 such that a subspace, which preserves the local structure better, can be learned without prior knowledge about the distance threshold. The experiments conducted on various datasets with several scenarios show the superiority of RUFSDR.