RipMC: RIPPER for Multiclass Classification

A major challenge in extending RIPPER for multiclass classification problems is the order of learning the classes. In this paper, RIPPER for Multiclass Classification (RipMC) is presented, which extends several aspects of RIPPER. In RipMC, all classes are initially given an equal opportunity with a Parallel Rule Learning (PRL) to generate their best rules in a global search, causing the rules in the decision list to be reordered, which improves performance in classifying new instances. Next, the most complex and costly class, which will be set as the default class in the subsequent execution of the algorithm, is identified according to a new measure called MaxDL. Finally, a new rule evaluation measure, namely LogLaplace, is presented for better pruning of the rules. The performance of the proposed algorithm and RIPPER is compared using 18 data sets. Experimental results show that RipMC significantly outperforms the original RIPPER.