A Nonparametric Feature Separability Measure and an Algorithm for Simulating Synthetic Feature Vectors

Measures that quantitatively reflect the separability between feature sets of two classes are required to identify the determinant features and select hyper-parameters of feature extraction algorithms, in binary classification paradigms. State-of-the-art separability measures look for equality of distribution parameters of the feature sets and do not linearly quantify the level of overlap between them. Reliable algorithms for generating synthetic feature sets with known levels of overlap are required to test and compare the performance of the separability measures. A measure of separability of features between two classes termed Thresholding-based Classification Error Estimate (TCEE) and an algorithm for generating synthetic feature vectors for testing the feature separability measures are proposed in this paper. Pearson’s correlation coefficient (PCC) of the Bhattacharyya distance (BD), Relative Entropy (RE), p-value of Rank-sum test, Jeffries-Matusita (JM) distance and TCEE with the percentage of overlaps on synthetic feature sets of two distinct classes are −0.6429, −0.6428, 0.3780, −0.9881, and 1. A high value of Pearson’s correlation with the percentage of overlap justifies that the TCEE can accurately measure separability of feature sets of two classes.