1 présentation

• 
  10h30 - 12h10

  Combinatorial Optimization for Trustworthy Machine Learning

  • Thibaut Vidal, prés., CIRRELT & SCALE-AI Chair in Data-Driven Supply Chains, Polytechnique Montréal

  The use of decision-support systems and machine learning algorithms in finance, medicine, and many other domains can profoundly impact human lives. Consequently, extensive efforts have been made to improve machine learning pipelines, e.g., making them more accurate, robust, and interpretable. In this tutorial, we discuss the synergy between combinatorial optimization and machine learning. We focus primarily on tree ensembles (including random forests and gradient boosting), a popular family of models with good empirical performance which is often used as a more transparent replacement to neural networks. We discuss the use of dynamic programming algorithms for simplifying tree ensembles. We also show that the search for counterfactual explanations (i.e., small and plausible perturbations of the input that modify the classification outcome) can be cast as a mixed integer program and efficiently solved. Finally, we use constraint programming to highlight privacy vulnerabilities in trained models. We conclude the talk with various research perspectives.