Prediction-Intervention Games and Invariant Sets in Causal Models
A recent preprint on arXiv (2605.16828) presents prediction-intervention games, a two-player Stackelberg framework. In this setup, a leader chooses a prediction function for a response variable Y from covariates derived from observational data, while a follower intervenes on certain covariates within a structural causal model to achieve their own goals. Although the leader is aware of the intervention targets, their understanding of the follower's objectives may be limited, complicating the task of finding an optimal strategy. To reduce performance loss, the leader can utilize predictions based on Y's causal parents or an invariant subset of covariates. The authors demonstrate that for two specific types of follower objectives, predictors leveraging the stable blanket—a defined invariant subset—perform at least as well as those relying on causal parents, and they establish an upper limit on the leader's potential performance loss.
Key facts
- arXiv preprint 2605.16828 introduces prediction-intervention games.
- The game is a two-player Stackelberg game.
- Leader chooses a prediction function for Y from covariates using observational data.
- Follower intervenes on some covariates in a structural causal model.
- Leader knows intervention targets but may have limited knowledge of follower's objective.
- Predictors based on causal parents or invariant subsets can reduce performance loss.
- Stable blanket is a specific invariant subset that outperforms causal parents for two follower objective classes.
- An upper bound on the leader's performance loss is provided.
Entities
Institutions
- arXiv