On Building Optimal Risk Scores through Mathematical Optimization
The widespread adoption of machine learning models in high-stakes domains underscores the need for interpretable and trustworthy predictive systems. Mathematical optimization has emerged as an effective tool for designing such models, offering explicit control over model structure and constraints. In this talk, we focus on the construction of risk scores—linear classifiers based on integer-weighted logistic regression—via convex mixed-integer non-linear optimization. In particular, we will leverage this convexity and explore an outer-approximation method.
Keywords: Interpretable Machine Learning; Risk Scores; Outer-Approximation