Exploiting the categorical reliability difference for binary classification

In binary pattern classification, the reliabilities of statistics obtained from the samples of the two categories are generally different. When the statistics are used for modeling a classifier, such reliability difference could impact the generalization performance. We formulate a disparity index to show the statistical disparity based on the generalized eigenvalue decomposition of the categorical moment matrices. It is shown that this disparity index can effectively indicate the reliability difference between the two categories. The obtained reliability difference is subsequently utilized to adjust the regularization term of a classifier for effective learning generalization. Our experiments based on 10 real-world benchmark data sets validate the effectiveness of the proposed method.