Prediction-Intervention Games and Invariant Sets

2026-05-19 · Source: stat.ML updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Mathematics & Computational Sciences · Depth: Expert, extended

Summary

This paper introduces "prediction-intervention games," a two-player Stackelberg game where a leader deploys a prediction function, and a follower reacts by intervening on covariates in a structural causal model (SCM) to optimize their own objective. The leader knows the intervention targets but may have limited knowledge of the follower's objective. The research demonstrates that predictors based on the "stable blanket," a specific invariant subset of covariates, are always superior or equal to those based on causal parents for two common follower objectives. The authors also upper-bound the leader's post-intervention risk by a worst-case risk over allowed interventions, providing conditions under which stable-blanket predictors are worst-case optimal. Practical strategies for learning these predictors, even when the causal graph is unknown, are proposed and tested on simulated and real-world data, including the Causal Chambers light tunnel.

Key takeaway

For AI Scientists and Research Scientists developing predictive models in dynamic environments, prioritizing predictors based on the stable blanket is crucial. This approach offers enhanced robustness against strategic follower interventions and can lead to lower post-intervention risk compared to relying solely on causal parents. You should integrate methods like stabilized classification, especially when the causal graph is uncertain, to ensure your models maintain performance and resist manipulation in real-world deployments.

Key insights

Stable blanket predictors consistently outperform or match causal parent predictors in prediction-intervention games.

Principles

• Invariant subsets improve prediction robustness.
• Stable blanket minimizes risk across environments.
• Worst-case risk bounds leader's post-intervention risk.

Method

The proposed method involves identifying invariant subsets, specifically the stable blanket, and training predictors on these subsets. For unknown graphs, it uses invariance tests and predictiveness scores to select optimal subsets or ensembles.

In practice

• Use stable blanket for robust predictions.
• Consider ensemble methods like stabilized classification.
• Test invariance with GCM for unknown causal graphs.

Topics

• Prediction-Intervention Games
• Structural Causal Models
• Stable Blanket
• Invariant Sets
• Distribution Generalization

Code references

• LinusKuehne/prediction-intervention-games
• juangamella/causal-chamber-package

Best for: AI Scientist, Research Scientist

Related on AIssential

• Partial Causal Structure Learning for Valid Selective Conformal Inference under Interventions — Selective conformal prediction under interventions can be robustly maintained by learning partial causal structures.
• The Impact of Dimensionality on the Stability of Node Embeddings — Node embedding stability varies with dimensionality, but optimal stability does not always align with peak performance.
• PAC-Bayesian Generalization Guarantees for Fairness on Stochastic and Deterministic Classifiers — PAC-Bayesian bounds can guarantee fairness for both stochastic and deterministic classifiers.
• Transferring Information Across Interventions in Causal Bayesian Optimization — GC-CBO improves causal Bayesian optimization by transferring knowledge across interventions via shared causal parameters.
• How Useful is Causal Invariance for Domain Adaptation in Finite-Sample Settings? — Causal invariance improves supervised domain adaptation in finite-sample settings when target-risk margins between candidate models are sufficiently large.
• Bounding the Predictive Space: How Topological AI Solves Catastrophic Forgetting Through… — Topological AI uses prime-indexed embedding anchors to prevent catastrophic forgetting in LLMs while retaining cognitive flexibility.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by stat.ML updates on arXiv.org.