Automatic Pruning of Rules Through Multi-objective Optimization—A Case Study with a Multi-objective Cultural Algorithm

Abstract

Classification algorithms create an overwhelmingly large number of rules, sometimes exceeding the number of data instances in the data set. Large number of rules hinders decision making. Therefore, there is a need for decreasing the rules before presenting it to the user. The process of removing unwanted rules is known in the data mining literature as rule pruning. In this pruning process, care should be taken to preserve the accuracy of the classifier. Mining compact classifiers with less number of rules that are also accurate and novel is a challenge. Thus, rule mining is a multi-objective optimization problem. Finding the best combination of subjective and objective metrics to present the user with compact, accurate and novel rules is the problem taken for study in the present study.