Consensus-based combining method for classifier ensembles

In this paper, a new method for combining an ensemble of classifiers, called Consensus-based Combining Method (CCM) is proposed and evaluated. As in most other combination methods, the outputs of multiple classifiers are weighted and summed together into a single final classification decision. However, unlike the other methods, CCM adjusts the weights iteratively after comparing all of the classifiers’ outputs. Ultimately, all the weights converge to a final set of weights, and the combined output reaches a consensus. The effectiveness of CCM is evaluated by comparing it with popular linear combination methods (majority voting, product, and average method). Experiments are conducted on 14 public data sets, and on a blog spam data set created by the authors. Experimental results show that CCM provides a significant improvement in classification accuracy over the product and average methods. Moreover, results show that the CCM’s classification accuracy is better than or comparable to that of majority voting.