Active Evidential Calibration of Binary SVM Classifiers

Abstract

Evidential calibration methods of binary classifiers improve upon probabilistic calibration methods by representing explicitly the calibration uncertainty due to the amount of training (labelled) data. This justified yet undesirable uncertainty can be reduced by adding training data, which are in general costly. Hence the need for strategies that, given a pool of unlabelled data, will point to interesting data to be labelled, i.e., to data inducing a drop in uncertainty greater than a random selection. Two such strategies are considered in this paper and applied to an ensemble of binary SVM classifiers on some classical binary classification datasets. Experimental results show the interest of the approach.