Abstract/Summary

Rule-based classifiers are considered more expressive, human readable and less prone to over-fitting compared with decision trees, especially when there is noise in the data. Furthermore, rule-based classifiers do not suffer from the replicated subtree problem as classifiers induced by top down induction of decision trees (also known as `Divide and Conquer'). This research explores some recent developments of a family of rule-based classifiers, the Prism family and more particular G-Prism-FB and G-Prism-DB algorithms, in terms of local discretisation methods used to induce rule terms for continuous data. The paper then proposes a new algorithm of the Prism family based on a combination of Gauss Probability Density Distribution (GPDD), InterQuartile Range (IQR) and data transformation methods. This new rule-based algorithm, termed G-Rules-IQR, is evaluated empirically and outperforms other members of the Prism family in execution time, accuracy and tentative accuracy.