In contemporary AI research, learning systems based on IF-THEN rule representations garner attention for their interpretability. The key objective of such rule sets is to achieve high discriminative power and interpretability. Existing state-of-the-art algorithms often prioritize predictive accuracy but fall short on quality metrics ensuring interpretability, such as coverage and parsimony. To address this, we propose CDPR, aiming to create highly accurate and interpretable rule sets for classification problems, marking the first attempt to establish such an approach. We introduce two algorithms rooted in submodular maximization, which not only provide provable guarantees on coverage but also yield rule sets that are both discriminative and parsimonious. Empirical results demonstrate that our learned rule sets achieve higher accuracy and interpretability, with more than a 2.5-fold improvement in average coverage rates compared to the next best algorithm.
Blogger's Review: The CDPR method presented in this paper strikes a commendable balance between interpretability and accuracy, especially with its remarkable improvement in coverage. This showcases the potential of the new algorithms and offers fresh perspectives for rule-based learning systems, warranting further research and application.