Boosting Approach for Score Level Fusion in Multimodal Biometrics Based on AUC Maximization

We investigate AdaBoost and bipartite version of RankBoost abilities to minimize AUC and its application for score level fusion in multimodal biometric systems. To do this, we customize two methods of weak learner training. Empirical results show comparable AUC for AdaBoost and RankBoost.B which previously was addressed theoretically. We demonstrate exhaustive results among state of the art classifiers and techniques, e.g., SVM, GMM and SUM rule in this area. AdaBoost and RankBoost.B achieve significant performance improvement compared to GMM and SUM rule, and the performance comparable to SVM. Besides empirical results, we show that, instead of adding a constant weak learner in order to maximize AUC using AdaBoost, instances could be weighted initially in each class inversely proportional to the number of instances in the corresponding classes.