Learning Optimal Classification Trees Using a Binary Linear Program Formulation
We provide a new formulation for the problem of learning the optimal classification tree of a given depth as a binary linear program. The running time of the existing Integer Linear programming (ILP) formulations increases dramatically with the size of data. We aim to circumvent this problem by making the formulation size largely independent from the training data size. We show that our formulation achieves better performance than existing formulations.
Sicco Verwer is assistant professor in machine learning with applications in cyber security and software engineering. He is best known for his work on learning state machines from trace data for which he received VENI and VIDI grants. Sicco is also interested in methods for declarative modelling of machine learning problems using mathematical solvers and making classifiers fair and robust.