A new decision fusion method in support vector machine ensemble

Abstract

In this paper, a new method of aggregating decisions in a multi-support vector machine (SVM) ensemble system is proposed. The evidence theory is introduced to reduce the uncertainty of decision-making. In the evidence theory, a practical problem is how to determine the basic probability assignments. Usually they are evaluated subjectively by experts in advance. However, they may be far from the optimal values. Furthermore, in some cases where there is no expert knowledge, especially for aggregation in an ensemble learning system, they could not be evaluated as such. Due to the natural relation between the evidence theory and the rough sets theory, rough sets methods are applied so as to determine the basic probability assignments. The merit of the rough set theory is that it does not need any priori knowledge. Afterwards, the decisions of bagged and boosted SVMs are combined respectively by the evidence theory. Experimental results show that the presented multi-SVM system gains better performance over the popular ensemble learning methods such as Bagging and Adaboost.M1.