An Integer Optimization Approach to Associative Classification

We aim to design classifiers that have the interpretability of association rules yet have predictive power on par with the top machine learning algorithms for classification. We propose a novel mixed integer optimization (MIO) approach called Ordered Rules for Classification (ORC) for this task. Our method has two parts. The first part mines a particular frontier of solutions in the space of rules, and we show that this frontier contains the best rules according to a variety of interestingness measures. The second part learns an optimal ranking for the rules to build a decision list classifier that is simple and insightful. We report empirical evidence using several different datasets to demonstrate the performance of this method.1 2