FEATURE SELECTION ALGORITHMS USING NON-REDUNDANT THRESHOLDED MEASURES

摘要

A new feature selection method, the threshold selection algorithm, is presented and compared with sequential selection and rejection algorithms. This algorithm assumes a measure of feature discrimination exists and provides a set of threshold parameters, associated with class pairs, which are dynamically variable. These thresholds provide a local as well as a global perspective to the problem of selection of feature subsets from a pool. Each threshold parameter provides an upper limit of class separation to be attained during feature selection;once this limit is reached, that class pair no longer affects the selection of features. Thus, the procedure maintains a global perspective by considering equally all class pairs which have not achieved their thresholds and in addition, particularly during the latter steps, focuses on the fewer local cases which have not been discriminated sufficiently.