Evidential joint calibration of binary SVM classifiers using logistic regression

Abstract

In a context of multiple classifiers, a calibration step based on logistic regression is usually used to independently transform each classifier output into a probability distribution, to be then able to combine them. This calibration has been recently refined, using the evidence theory, to better handle uncertainties. In this paper, we propose to use this logistic-based calibration in a multivariable scenario, i.e., to consider jointly all the outputs returned by the classifiers, and to extend this approach to the evidential framework. Our evidential approach was tested on generated and real datasets and presents several advantages over the probabilistic version.