Transformation from Local Accuracy to Classification Confidence

Local accuracy, which represents the accuracy of a base classifier for an input pattern, is one kind of valuable information used in the classifier combination methods. However, the only existing classifier combination method which uses local accuracy, namely dynamic classifier selection, can not take full advantage of the information from the base classifiers. In dynamic classifier selection, the final classification result is determined by the base classifier with the highest local accuracy, and the local accuracies of the other base classifiers are neglected. In this paper, a method of transforming local accuracy into classification confidence is proposed, in which the confidence value corresponding to the class outputted by a base classifier is proportional to the local accuracy of the base classifier, and the confidence values corresponding to the other classes are assumed to be equal. After the transformation, multiple classifier systems can be designed with measurement level classifier fusion methods such as linear combination. Compared with dynamic classifier combination, the classifier fusion methods, which make decisions by integrating the classification confidences of all the base classifiers, can use more information and achieve a higher classification rate. To evaluate this approach, a lot of experiments are conducted on six large data sets selected from the ELENA, UCI, and DELVE databases. The experimental results show that the approach outperforms dynamic classifier selection by 0.2% to 13.6% in classification rate.