Local Selective Voting

We propose a technique of localized selective voting of weak classifiers. This technique identifies local regions having similar characteristics and then uses the selective votes of each local expert to describe the relationship between the data characteristics and the target class. The algorithms that are locally used for building the ensemble are tested in the local set and if they have statistically worse accuracy than the most accurate algorithm, they do not participate to the final decision of the ensemble. Our experiment for several UCI datasets shows that the proposed combining method outperforms other combining methods we tried as well as any base classifier.