Unsupervised Feature Selection

Abstract

Abstract. The problem of feature selection has long been an active research topic within statistics and pattern recognition. So far, most methods of feature selection focus on supervised data where class information is available. For unsupervised data, the related methods of feature selection are few. The presented article demonstrates a way of unsupervised feature selection, which is a two-level filter model removing the redundant and irrelevant features, respectively. The redundant features are eliminated using any clustering algorithm, and a new method is proposed to remove the irrelevant features: first rank the features according to their relevance to cluster and then a subset of relevant features is selected using the Fuzzy Feature Evaluation Index (FFEI) with some changes and extensions. The experimental results have shown the effectiveness of the proposed method for high-dimensional data. Our major contributions are: (1) to present a new hierarchical filter method for unsupervised feature selection; (2) to propose a new algorithm for removing the irrelevant features; (3) to extend the FFEI, and present a method for calculating the approximate weight of feature in FFEI, which improves the efficiency and robustness of the method.