A Hybrid Method of Unsupervised Feature Selection Based on Ranking

Abstract

Feature selection is a key problem to pattern recognition. So far, most methods of feature selection focus on sample data where class information is available. For sample data without class labels, however, the related methods for feature selection are few. This paper proposes a new way of unsupervised feature selection. Our method is a hybrid approach based on ranking the features according to their relevance to clustering using a new ranking index which belongs to exponential entropy. Firstly a candidate feature subset is selected using a modified fuzzy feature evaluation index (FFEI) with a new method to calculate the feature weight, which makes the algorithm to be robust and independent of domain knowledge. Then a wrapper method is used to select compact feature subset from the candidate feature set based on the clustering performance. Experimental results on benchmark data sets indicate the effectiveness of the proposed method