Combining Answers of Sub-classifiers in the Bagging-Feature Ensembles

摘要

Improving classification performance of learning systems can be achieved by constructing multiple classifiers which include sets of sub classifiers, whose individual predictions are combined to classify new objects. The diversification of sub-classifiers is one of necessary conditions for improving the classification accuracy. To obtain more diverse sub-classifiers we extend the bagging approach by integrating sampling different distributions of learning examples with selecting multiple subsets of features. We summarize results of our experiments on studying the usefulness of different feature selection techniques in this extension. The main aim of the paper is to examine the use of three methods for aggregating predictions of sub-classifiers in the extended bagging classifier. Our experimental results show that the extended classifier, with a dynamic choice of answers instead of a simple voting aggregation rule, is more accurate than standard bagging.