A Feature Selection Algorithm Based on SVM Average Distance

Feature selection is a very important part for datamining, machinery learning and pattern recognition. Distance plays a vital role in Support Vector Machines (SVM) theory. Relief-F algorithm solves feature redundancy well but doesn´t guarantee the maximum distance. To overcome this problem, a feature subset selection algorithm is proposed which takes SVM average distance as estimation rule and sequential forward selection as search strategy. Using public data set acquired from UCI, this algorithm is compared with the Relief-F. The results show that the recognition rate is higher than Relief-F with smaller selected features under computation amount tolerant conditions.