ADIGen: New Framework for Debiased, Invariant Counterfactual Generation

Generating accurate counterfactual outcomes is critical for decision-making under complex interventions, but existing generative models often suffer from unstable estimation, poor generalization across environments, and bias from nuisance model misspecification. A new paper by Raphael C Kim, Jingsen Zhu, Ramin Zabih, and Michele Santacatterina introduces ADIGen (Automatic, Debiased, and Invariant Counterfactual Generation), a framework designed to overcome these challenges. ADIGen employs Riesz regression for efficient density-ratio estimation, causal invariance principles to ensure better performance under distribution shifts, and orthogonal statistical learning to provide doubly robust guarantees against misspecified nuisance models. The authors derive excess-risk bounds confirming that ADIGen controls counterfactual risk under general interventions, including high-dimensional outcomes.

This approach marks a significant advance in causal machine learning by offering theoretical guarantees where previous methods relied on brittle assumptions. The combination of invariant learning and doubly robust estimation allows ADIGen to generate counterfactual samples that remain reliable even when the training environment differs from deployment conditions. The framework automatically handles both the intervention and outcome dimensions without needing manual tuning or strong parametric assumptions. While the paper is theoretical, the methodology is directly applicable to areas like personalized medicine, economics, and policy evaluation where robust counterfactual reasoning is essential. The authors demonstrate excess-risk bounds with a product-bias nuisance remainder and an invariant risk bound across environments, solidifying ADIGen as a principled solution for automatic counterfactual generation.