Abstract. We consider the problem of building Boolean rule sets in disjunctive normal form (DNF), an interpretable model for binary classification, subject to fairness constraints. We formulate the problem as an integer program that maximizes classification accuracy with explicit constraints on equality of opportunity and equalized odds metrics. A column generation framework is used to efficiently search over exponentially many possible rules, eliminating the need for heuristic rule mining. Compared to other interpretable machine learning algorithms, our method produces interpretable classifiers that have superior performance with respect to the fairness metric.

Joint work with Connor Lawless

Bio: Oktay Gunluk joined the School of Operations Research and Information Engineering faculty in January 2020. Before joining Cornell, he was the manager of the Mathematical Optimization and Algorithms group at IBM Research. He has also spent three years as a researcher in the Operations Research group in AT&T Labs. At both of these industrial labs, in addition to basic research in mathematical optimization, he has worked on various large-scale applied optimization projects for internal and external customers. His main research interests are related to theoretical and computational aspects of discrete optimization problems, mainly in the area of integer programing. In particular, his main body of work is in the area of cutting planes for mixed-integer sets. Some of his recent work focuses on developing integer programming-based approaches to classification problems in machine learning. He has B.S./M.S. degrees in Industrial Engineering from Boğaziçi University, and M.S./Ph.D. degrees in Operations Research) from Columbia University.