A Proposal of Ensemble Learning \pdi-BoostingG" and its Evaluation

Abstract

Recently, we have been increasing interest in ensemble learning. In particular, Boosting is a useful learning method in which multiple classifiers construct multiple layers with maintaining mutual dependencies, and its expressive ability is high. We have proposed pdi-Boosting, which virtually generates interpolated data and adds them to the training data to improve the accuracy of classifiers. However, in this method, the area for generating the additional fuzzy rule is not directly related to the misidentification data and virtual data. Therefore, the discrimination rate is not high. In this paper, we propose a new \pdi-BoostingG" that also regularizes the additional region G of fuzzy rules. In pdi-BoostingG, fuzzy rules are additionally generated around the misidentification data, and virtual data is generated in this area G. In addition, the individual learning of the additional fuzzy rule and the whole learning of the whole space are alternately learned to form the multi-layered structure. As a result, pdi-BoostingG improves the discrimination rate and robustness. In addition, since virtual data and fuzzy rules are inherited between multiple layers, deep inference of fuzzy rules is realized. We formulate here how to generate additional fuzzy rules and virtual data, and how to inherit whole fuzzy rules and virtual data between layers. In addition, we compare the discrimination rate of pdi-BoostingG with other methods in order to discuss its usefulness.