An Unsupervised Learning Approach for Case-Based Classifier Systems

Case-Based Classifier Systems obtain low accuracies on generalisation and higher waste on CPU time when the class distribution space is not well defined. This paper presents the Mean Sphere and the Mean K-Means approach based on unsupervised learning to improve the CPU time and to improve or maintain the accuracy. We use clustering in an unsupervised way to decide what is the representational space of each class. The concept of clustering is introduced in two levels. The first level clusters the training data into spheres, obtaining one sphere for each class. The second level consists of clustering the spheres in order to detect the behaviour of the elements present in the sphere. In this level two policies are applied, Mean Sphere and Mean K-Means approaches. Experiments using different domains, most of them from the UCI repository, show that the CPU time is considerably decremented while maintaining, and sometimes improving, the accuracy of the system.