Decision-Focused Learning for the Masses With Applications to Public Health

Abstract

In this thesis, I aim to better understand how to best train Machine Learning models for decision-making under uncertainty. Specifically, I focus on the ``Predict-then-Optimize'' framework, in which uncertain quantities are predicted using ML models, and decisions are made by solving optimization problems parameterized by these predictions. While past Decision-Focused Learning (DFL) methods show that optimizing directly for decision quality leads to improved outcomes, existing approaches typically require extensive manual effort—such as designing differentiable surrogate optimization tasks—limiting their applicability to arbitrary problems. Moreover, evaluating the real-world impact of these models in resource-constrained settings poses additional challenges. This thesis, therefore, addresses the central question: Can we create generalizable methods to train and evaluate decision-focused learning models on arbitrary decision-making tasks so that DFL can be more practically useful? In response, I introduce methods that distill task-specific decision-making structures into a learned, differentiable ``decision loss,'' eliminating the need for handcrafted surrogates. I first propose Locally Optimized Decision Losses (LODLs), demonstrating their improved performance across multiple domains. I then extend this approach via Efficient Global Losses (EGLs), significantly enhancing generalization, efficiency, and theoretical robustness. Additionally, I develop rigorous statistical estimators for accurately evaluating DFL models in resource allocation scenarios. Applying these estimators to real-world randomized control trials reveals insights previously hidden by existing methods. Collectively, these contributions establish a broadly applicable, reliable framework for decision-focused learning, making DFL more practically viable across diverse decision-making contexts.