Machine Learning and Data Mining in Pattern Recognition, Posters pages:216-230
International Conference on Machine Learning and Data Mining in Pattern Recognition edition:5 location:Leipzig date:17-20 July 2007
Fusion of an ensemble of multiple classifiers can result in more accurate classification results than each of the individual members of the ensemble. This fusion is often based on fixed combination rules, which are almost never optimal. A better performance can be achieved by learning an adequate way to combine the classifiers in the ensemble.
In this work a trainable classifier fusion method is proposed, based on the Dempster-Shafer theory of evidence. This method takes into account both information about the accuracy of the classifiers and the information about their "most typical outputs", which are processed by the orthogonal sum rule and the discounting operation from Dempster-Shafer theory. This classifier fusion method is evaluated on four data sets from the UCI repository, comparing it with nine other fusion methods. Our results show that this method is superior to the other methods which were tested on all four data sets.