CCC: Classifier combination via classifier

Abstract

The combination of classifier has long been proposed as a method to improve the accuracy achieved in isolation by a single classifier. Most of the extant works focus on how to generate a group of "good" base classifiers, such as AdaBoost and Bagging. We are interested in the method of combining multiple classifiers. In contrast to such popular used method as vote, we regard the classifier combination problem as a classification problem. From the perspective of pattern recognition, the base classifiers can also be regarded as a feature extraction method. In theory, any of classifiers can be used to combine the base classifiers as long as they are able to treat the outputs of base classifiers. More generally, the combination model is also able to deal with other machine learning problems including cluster and regression task, that is Learner Combination via Learner(LCL) model. A large empirical study shows that, comparing with majority vote(Bagging) and weighted majority vote(AdaBoost), CCC can significantly improve the performance. © 2011 Springer-Verlag.