Information fusion of agent based heterogeneous multi-classifiers

Traditional technology of classifier fusion cannot make full use of the characteristics of heterogeneous classifiers to deal with various problems. This work suggests a new technology of information fusion using multiple agents, each of which uses a quite different classification algorithm such as decision tree algorithm, simple Naive Bayes algorithm and the newly emerging classification algorithm based on atomic association rules. Information fusion of these heterogeneous multi-classifiers is based on the classifier behavior, properties of training dataset and the instance to be classified. The proposed technology has following advantages: (1) high classification accuracy, (2) no need of fusion training, and (3) fast learning and prediction. The experimental results on 10 UCI standard datasets show that accuracy of the proposed fusion technology is noticeably higher than that of traditional voting method.