New study proves stable blanket beats causal parents in ML games

Researchers from the University of Copenhagen and ETH Zurich (Kühne, Schur, Peters) have published a paper on arXiv (2605.16828) tackling a fundamental problem in robust ML: what happens when an adversary can intervene on your input features after you've deployed a predictor? They formalize a two-player Stackelberg game where a leader fits a prediction function for Y from covariates X, and a follower then chooses an intervention on some covariates to maximize their own objective. The leader knows which variables can be intervened on but not the follower's exact objective. Finding an optimal strategy for the leader is NP-hard in general, but the authors show a surprisingly elegant solution.

Their key contribution is proving that predictors built on the stable blanket — a specific invariant subset of covariates that is maximally predictive under all possible interventions — always match or outperform predictors based on the causal parents of Y. They derive upper bounds on the leader's post-intervention risk by analyzing worst-case risk over allowed interventions. The paper gives sufficient conditions under which stable-blanket predictors are worst-case optimal, and shows by counterexamples that these conditions cannot be dropped. Experiments on both synthetic data and real-world datasets (e.g., from genomics) confirm the theoretical results. The work bridges causal inference, game theory, and robust machine learning, offering practical guidance for deploying predictors in adversarial environments.