A Possibilistic Rule-Based Classifier

Rule induction algorithms have gained a high popularity among machine learning techniques due to the “intelligibility” of their output, when compared to other “black-box” classification methods. However, they suffer from two main drawbacks when classifying test examples: i) the multiple classification problem when many rules cover an example and are associated with different classes, and ii) the choice of a default class, which concerns the non-covering case. In this paper we propose a family of Possibilistic Rule-based Classifiers (PRCs) to deal with such problems which are an extension and a modification of the Frank and Witten’ PART algorithm. The PRCs keep the same rule learning step as PART, but differ in other respects. In particular, the PRCs learn fuzzy rules instead of crisp rules, consider weighted rules at deduction time in an unordered manner instead of rule lists. They also reduce the number of examples not covered by any rule, using a fuzzy rule set with large supports. The experiments reported show that the PRCs lead to improve the accuracy of the classical PART algorithm.