Building Ensembles with Classifier Selection Using Self-Organizing Maps

Improving the performance of supervised classification methods is a subject of many literature works. An efficient strategy is the adoption of an ensemble of classifiers to divide the classification problem. In ensembles with classifier selection, there is no fusion of the classifiers decisions. A particular classifier is selected according to the input data instead. In this paper, well-known clustering methods based on self-organizing structures are used to implement ensembles with classifier selection. The self-organizing structures are used to detect the topological structure of data and help to divide the problem into smaller and easier sub-problems to solve. Experiments with different datasets show that the use of clustering methods to perform the classifier selection can contribute to split the problem and improve the classification accuracy compared to some traditional strategies. Additionally, the results encourage the development of more research to find out other ways to split problems using data clustering techniques.