A Dual Perspective on Decision Focused Learning

Predict-then-optimize (PtO) pipelines use predictions as inputs to a downstream optimizer that produces decisions. Decision-Focused Learning (DFL) trains the predictor inside this pipeline to improve the quality of those decisions, not just prediction accuracy. Most DFL approaches do so by differentiating through the optimizer or by designing tight task-specific surrogates—both of which demand frequent solver calls and drive up training cost. We introduce a dual-guided DFL method that preserves decision alignment while sharply reducing solver-in-the-loop cost. The key idea is to solve the downstream problem only periodically to refresh dual variables and, between refreshes, train on dual-adjusted targets using simple surrogate losses. As refreshes become less frequent, the training loop’s cost approaches standard supervised learning while maintaining high-quality decisions.