Performance Loss of Ignoring Correlation in Binary Classification under Gaussian Model: A Numerical Perspective M. Syamsuddin Wisnubroto, Sumardi, and Nanang Susyanto
Department of Mathematics, Universitas Gadjah Mada
Abstract
Given two or more classifiers, researchers usually assume independence between classifiers to simplify binary classification tasks. However, in fact, two classifiers are dependent because they often process the same input with a different method. In this paper, we investigate numerically the effect of ignoring dependence between classifiers when the distribution of each class is Gaussian. We compute the performance loss of assuming independence between classifiers, which is defined by the difference between the True Positive Rate of the model with dependence and the model without dependence for any fixed False Positive Rate. In particular, we emphasize analyzing the relation between the correlation of each class with the performance loss under the independence assumption.
Keywords: Likelihood Ratio Dependent Classifiers Performace Loss Binary Classification
