Decision-Focused Learning with Directional Gradients

Michael Huang, Vishal Gupta

We propose a novel family of decision-aware surrogate losses, called Perturbation Gradient (PG) losses, for the predict-then-optimize framework. The key idea is to connect the expe…