Feature Usability Index and Optimal Feature Subset Selection

usability index is introduced here as a measure for evaluat- ing classification efficacy of features. It is defined using measures of homogeneity, class specificity, and error in decision making. Homogeneity measures the extent of outlying observations, class specificity assesses the separation between distributions of differ- ent labeled classes, and error in decision making is computed using overlap in posteriori decision boundary. This is followed by feature ranking and optimal feature subset selection through ordering of features based on feature usability index and involves a complexity of O(DlogD) for D features. The results validating classifier in- dependent feature ranking and optimal feature subset selection are also presented aong with a comparative analysis using χ 2 statistics for feature selection. General Terms Pattern Reconition, Machine Intelligence, Data Mining