RBF ensemble based on reduction of DAG structure

Binary classifiers are grouped into an ensemble to solve multi-class problems. One of proposed ensemble structure is a directed acyclic graph. In this structure, a classifier is created for each pair of classes. The number of classifiers can be reduced if groups of classes will be separated instead of individual classes. The proposed method is based on the similarity of classes defined as a distance between classes. For near classes the structure of DAG stays immutable. For the distant classes more than one is separated with a single classifier. In this paper, the proposed method is tested in variants based on various metrics. For the tests, several datasets from UCI repository was used and the results were compared with published works. The tests proved that grouping of radial basis functions into such ensemble reduces the classification cost and the recognition accuracy is not reduced significantly.