Efficient Learning of Interpretable Classification Rules

Machine learning has become omnipresent with applications in various safety-critical domains such as medical, law, and transportation. In these domains, high-stake decisions provided by machine necessitate researchers to design interpretable models, where the prediction is understandable a human. learning, rule-based classifiers are particularly effective representing decision boundary through set of rules comprising input features. Examples include trees, lists, sets. The interpretability general related size rules, smaller considered more interpretable. To learn classifier, brute-force direct approach consider an optimization problem that tries smallest classification rule close maximum accuracy. This computationally intractable due its combinatorial nature thus, not scalable large datasets. this end, paper we study triangular relationship among accuracy, interpretability, scalability classifiers. contribution framework IMLI, based on satisfiability (MaxSAT) for synthesizing expressible proposition logic. IMLI considers joint objective function optimize accuracy learns optimal solving appropriately designed MaxSAT query. Despite progress last decade, straightforward MaxSAT-based solution cannot scale practical datasets containing thousands millions samples. Therefore, incorporate efficient incremental technique inside formulation integrating mini-batch iterative rule-learning. resulting classifier iteratively covering training data, wherein each iteration, it solves sequence queries corresponding mini-batch. our experiments, achieves best balance scalability. For instance, attains competitive w.r.t. existing demonstrates impressive both non-interpretable fail. As application, deploy popular lists source code available at https://github.com/meelgroup/mlic.