Ensemble unsupervised feature selection based on permutation and R-value

Selecting the informative features from the high dimensional data can improve the performance of the classification and get a deep understanding of the problems. A non-problem related feature contains little information and has little influence on the data distribution. By permuting the feature and calculating the data distribution difference, how much information the feature contains could be measured. In this paper, we propose an unsupervised feature selection method (EUFSPR), which combines the ensemble technique, clustering, permutation and data distribution evaluation techniques to measure the feature importance. Clustering is adopted to get the sample groups and the data distribution is evaluated by the overlapping areas. Eight gene expression microarray datasets are utilized to demonstrate the effectiveness of the proposed method over the unsupervised feature selection methods and supervised feature selection methods.