Multiple Classifier Systems for Protein Function Prediction

Classification problems have gained increasing attention in data mining fields. Multiple classifier systems have been actually proved to reach higher classification accuracy and reliability than single classifier, and can be designed by two approaches: modular and ensemble. Our concern is to adopt ensemble, where each component classifier independently and essentially performs the same classification task, and then the outputs are combined as the final output with special fusion rule, thus multiple classifier systems fusion belongs to the decision-level fusion. It is significant to examine the fusion strategy. In this paper, we studied a fusion method based on different classifiers with overproduction and selection strategy, and applied it to the prediction for a particular type of protein: G-protein coupled receptors, which is very important as they can change a cell's behavior by transmitting messages from the cell's exterior to its interior. The performance of the multiple classifier system is evaluated and some comparable experiment results are obtained.