Learning Optimal Decision Trees Using MaxSAT

We present a Combinatorial Optimization approach based on Maximum Satisfiability technology to compute Minimum Pure Decision Trees (MPDTs) for the sake of interpretability. We show that our approach outperforms clearly in terms of runtime previous approaches to compute MPDTs. We additionally show that these MPDTs can outperform on average the DT classifiers generated with sklearn in terms of accuracy. Therefore, our approach tackles favourably the challenge of balancing interpretability and accuracy.