Improving Modular Classification Rule Induction with G-Prism Using Dynamic Rule Term Boundaries

Abstract

Modular classification rule induction for predictive analytics is an alternative and expressive approach to rule induction as opposed to decision tree based classifiers. Prism classifiers achieve a similar classification accuracy compared with decision trees, but tend to overfit less, especially if there is noise in the data. This paper describes the development of a new member of the Prism family, the G-Prism classifier, which improves the classification performance of the classifier. G-Prism is different compared with the remaining members of the Prism family as it follows a different rule term induction strategy. G-Prism’s rule term induction strategy is based on Gauss Probability Density Distribution (GPDD) of target classes rather than simple binary splits (local discretisation). Two versions of G-Prism have been developed, one uses fixed boundaries to build rule terms from GPDD and the other uses dynamic rule term boundaries. Both versions have been compared empirically against Prism on 11 datasets using various evaluation metrics. The results show that in most cases both versions of G-Prism, especially G-Prism with dynamic boundaries, achieve a better classification performance compared with Prism.