<ul><li>Generative classifiers are typically learned using closed-form procedures that maximize data fitting scores, but are not directly linked to supervised classification metrics.</li><li>To address this limitation, a learning procedure called risk-based calibration (RC) is proposed, which adjusts the joint probability distribution of the classifier according to the 0-1 loss in training samples.</li><li>RC reinforces data statistics associated with true classes and weakens those of incorrect classes, progressively improving the classifier's training error.</li><li>Experimental results on 20 datasets show that RC outperforms closed-form learning procedures in terms of training error and generalization error, bridging the gap between traditional generative approaches and performance-guided learning procedures.</li></ul>

Risk-based Calibration for Generative Classifiers

Discover more