Mining interpretable rules with MCRM: A novel rule mining algorithm with inherent feature selection and discretization

This paper presents MCRM, a new meta-heuristic rule learning algorithm based on the Pittsburgh approach. MCRM extracts all rules simultaneously with a single execution of a meta-heuristic algorithm. The proposed algorithm uses the artificial bee colony (ABC) as a meta-heuristic optimization algorithm. In the MCRM algorithm, classification rules are extracted as a result of a multivariate discretization process and the task of the ABC optimization algorithm is to guide the discretization process. The introduced structure for rule learning guarantees that the extracted rules are mutually exclusive. It also allows the algorithm to ignore the less important features while learning the rules and perform the feature selection operation implicitly. Also, the algorithm uses a special post-pruning step for pruning the final rules, which helps to increase the interpretability of the final rules by reducing the number and length of the rules. The proposed algorithm is compared with 15 other classification algorithms on 32 datasets from the OpenML website and the UCI machine learning repository. The experimental results show the promising performance of the proposed algorithm compared to other classifiers applied in the experiments.