A study on similarity measures for the feature selection method OFFSS

Feature subset selection has been playing a very important role on data-mining and pattern recognition. OFFSS (optimal fuzzy-valued feature subset selection) is a new feature selection method that selects an optimal feature subset by considering both the overall overlapping degree between two classes of examples and the size of feature subset. In comparison with other methods such as OFEI, FQI and MIFS, OFFSS has no significant difference in training accuracy of the selected feature subset but has much less computational complexity. Since, the OFFSS algorithm is dependent of a similarity measure so that different similarity measures may lead to different feature subsets to be selected. We study the impact of numerical results of similarity measures on the results of OFFSS for the same dataset. Based on triangle membership functions, we demonstrate the relationship among threshold, feature subset, and classification accuracy that are produced by OFFSS using three classes of similarity measures respectively. Then one similarity measure is found for selecting less number of features.