Ant System and Weighted Voting Method for Multiple Classifier Systems

Combining multiple classifiers is considered as a general solution for classification tasks. However, there are two problems in combining multiple classifiers: constructing a diverse classifier ensemble; and, constructing an appropriate combiner. In this study, an improved multiple classifier combination scheme is propose. A diverse classifier ensemble is constructed by training them with different feature set partitions. The ant system-based algorithm is used to form the optimal feature set partitions. Weighted voting is used to combine the classifiers’ outputs by considering the strength of the classifiers prior to voting. Experiments were carried out using k-NN ensembles on benchmark datasets from the University of California, Irvine, to evaluate the credibility of the proposed method. Experimental results showed that the proposed method has successfully constructed better k-NN ensembles. Further more the proposed method can be used to develop other multiple classifier systems.