Prediction-Intervention Games and Invariant Sets
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "Prediction-Intervention Games and Invariant Sets".
Jane: Here's the gist of the paper. We consider the following two-player game: using observational data, the leader chooses a prediction function for a response variable Y from given covariates The single most informative sentence about the paper's subject and finding is that predictors based on the stable blanket are always better or as good as those based on the causal parents Prediction-Intervention Games The paper formalizes the problem as a causal version of a Stackelberg game, which it calls a prediction-intervention game, where the leader provides a prediction function f and the follower reacts by choosing an intervention e to maximize their objective The leader seeks to find argminf Rlead e(f) (f), while the follower selects e in argmaxe V foll e(f) (f) This setup is a special case of a Stackelberg game.
Tom: Invariant Sets and Predictors The paper introduces the concept of an invariant subset S, defined as a subset where Y ⊥⊥ E XS under P, or equivalently, if Y XS is the same under all Pe, e ∈ E The stable blanket SB(Y) is defined as a set satisfying PA(Y) ⊆ SB(Y) ⊆ MB(Y), which captures as much information about Y as possible while excluding irrelevant covariates.
Jane: Optimality of the Stable Blanket Theorem 3 states that for two common follower objectives, either V foll e(f) = Re(f) or V foll e(f) = Ee
α + β f(X) − ce(X): for α, β ∈ R and a family of measurable functions on X that do not depend on f, then sup e∈E∗(fSB(Y)) Re(fSB(Y)) ≤ sup e∈E∗(fPA(Y)) Re(fPA(Y)) This proves that predictors based on the stable blanket are better or as good as those based on the causal parents under these conditions.
Tom: Worst-Case Distribution Generalization The worst-case risk over the possible interventions upper-bounds the leader’s post-intervention risk and does not require knowledge of the follower’s objective The bound (4) shows that sup e∈E∗(f) Re(f) ≤ sup e∈E Re(f), which connects prediction-intervention games to distribution generalization.
Paper summary: Tom: So we're looking at this paper now titled "Prediction-Intervention Games and Invariant Sets." Basically it sets up a two-player game where the leader picks a prediction function f from some data to guess a response variable Y, and then a follower reacts by choosing an intervention on some variables in the underlying causal model to try and maximize their own goal.
Jane: That sounds complex because the leader only knows what they can predict, but they don't know what the follower is actually trying to do, right? This paper formalizes that as a prediction-intervention game, which is just a special kind of Stackelberg game.
Lu: The core idea here is finding an optimal strategy for the leader, which we call minimizing the risk Rlead e (f) over all possible interventions e in E at the leader's level <ref:2605.16828#pg3>. To avoid terrible performance loss, the leader can base their prediction on the causal parents of Y, or more broadly on some invariant subset of covariates <ref:2605.16828#pg1>.
Meng: If we think about deployment, what does this mean for us when we're shipping models? The abstract says that finding an optimal strategy for the leader is generally really hard to figure out <ref:2605.16828#pg3>. So the paper suggests that instead of trying to find that perfect leader strategy, we might base our predictions on something simpler, like an invariant subset or the causal parents.
Lalam: The specific finding they prove is pretty strong: for two common classes of follower objectives, predictors based on the stable blanket are always better or as good as those based on the causal parents <ref:2605.16828#pg1>. This suggests that this stable blanket approach is a very robust way to predict Y.
Tom: Wait, "better or as good" sounds like a strong claim. So what's this stable blanket thing they're talking about? Jane, can you explain what that invariant set is supposed to capture?
Jane: Well, the paper defines an invariant subset S as a set where Y is conditionally independent of the interventions E given that subset S <ref:2605.16828#pg1>. It’s essentially a group of variables where the relationship between Y and those variables doesn't change no matter what intervention happens on other variables <ref:2605.16828#pg1>.
Paper summary: Lu: They define the stable blanket SB(Y) as a set that satisfies two conditions: first, the distribution of Y given that subset S is contained within SB(Y), and second, it captures as much information about Y as possible while excluding irrelevant covariates <ref:2605.16828#pg1>. It’s like finding the tightest group of features that actually matter for predicting Y.
Meng: So if we think about running a model in production, what's the practical difference between basing our prediction on the causal parents versus this stable blanket? Does one of them introduce more risk?
Lalam: Theorem three proves that under two common follower objectives, either the follower maximizes their objective by just using the prediction function f itself or by using a linear combination involving some functions on X <ref:2605.16828#pg1>. And in those cases, predictors based on the stable blanket yield equal expected utilities across all environments <ref:2605.16828#pg3>.
Tom: Equal expected utilities sounds good, but we need to be careful about the conditions they put on that. The paper mentions a worst-case risk over possible interventions which upper bounds the leader’s post-intervention risk <ref:2605.16828#pg4>. That's a nice connection to distribution generalization.
Jane: It connects this game setup to how we generalize models under different data distributions, showing that the worst-case risk over all possible interventions is upper bounded by the leader's risk in a specific way <ref:2605.16828#pg4>. The theorem they prove shows that predictors based on the stable blanket solve this worst-case risk problem better than those based on causal parents under certain assumptions <ref:2605.16828#pg3>.
Lu: The method they propose for actually finding these predictors is called the IMP approach, which stands for invariant most-predictive <ref:2605.16828#pg4>. It involves searching for subsets S that are both invariant on the intervention set and predictive of Y <ref:2605.16828#pg4>. They even suggest stabilized regression and classification methods to handle errors in that invariance test, which is a practical consideration for us.
Meng: If we want to deploy this, what's the cost of finding this invariant most-predictive subset S? Is it computationally expensive for our current model sizes?
Paper summary: Lalam: The implementation details are laid out in sections E.four and F.five showing they use multilayer perceptrons for estimation and optimization under specific objectives <ref:2605.16828#pg3>. It's not just a theory; they show how to actually build the predictor on hardware that we use today.
Tom: So, to wrap up this first look of "Prediction-Intervention Games and Invariant Sets," it’s about formalizing how leaders choose predictors in games where the follower can actively mess with the data or model structure. The key finding is that using predictors based on the stable blanket is always as good as, or better than, using predictors based on causal parents under specific conditions <ref:2605.16828#pg1>.
Jane: It really matters for us because it gives us a more robust way to select features when we don't have perfect knowledge of the entire causal structure <ref:2605.16828#pg4>. It moves beyond just looking at direct causal links to finding sets of variables that are truly invariant under the influence of potential interventions.
Lu: The authors conclude that this subset view is arguably more interpretable and transparent than using complex function-based invariance criteria like IRM <ref:2605.16828#pg1>. It combines feature selection with standard supervised learning, which simplifies auditing and makes individual predictions easier to explain for an engineer <ref:2605.16828#pg4>.
Meng: So, the implication is that instead of getting bogged down in complex functional constraints, we should focus on finding these invariant subsets S that are also predictive of Y <ref:2605.16828#pg4>. It makes sense for production systems because it’s more auditable.
Lalam: They also note that the results rely on the assumption of strong interventions, and they leave the finite-sample analysis for future work <ref:2605.16828#pg4>. That's a key caveat we need to keep in mind when applying this to real traffic data.
Tom: It sounds like a solid piece of theoretical work that bridges causal inference with model selection, and it gives us a concrete way to improve prediction robustness <ref:2605.16828#pg1>. We'll be keeping an eye on how we can use these stable blanket predictors in our next training cycle.
Conclusion: Tom: So we're wrapping up our look at "Prediction-Intervention Games and Invariant Sets." It’s basically about formalizing this whole game where one player predicts things, and another player tries to mess with the system to get their own goal.
Jane: Right, so it’s taking a prediction problem and turning it into a game between two people. The authors are looking at how the leader makes choices when they don't know exactly what the follower is doing.
Lu: The main thing they establish is that using this concept of a stable blanket for predictors always gives you results that are as good as, or better than, basing your prediction on just the direct causal parents.
Meng: So, in plain terms for someone running a model on production data, it means you don't need to perfectly map out every single variable relationship if you use this stable blanket method.
Lalam: The authors show that these predictors are robust because they ignore irrelevant variables that don't actually matter for the outcome under different conditions.
Tom: It sounds like the paper is arguing that focusing on these invariant sets instead of just the most obvious causal links is a way to get a more stable prediction.
Jane: Exactly, it’s about finding those specific groups of features where things stay consistent no matter what other stuff changes around them.
Lu: They propose a method called IMP, which means invariant most-predictive, as their way to actually find these subsets in the data.
Meng: That makes sense because if we have to pick a predictor for deployment, you want something that doesn't break easily when the environment shifts slightly.
Lalam: And they did test this out on some real data from a light tunnel, showing that this stable blanket approach works even when the causal graph assumptions are not perfect.
Tom: It seems like the authors really believe this subset view is more practical for building reliable prediction systems than relying only on those direct causal parents.
Jane: That’s what they conclude, suggesting it’s easier to audit and explain because you're just looking at a specific set of features that are truly important.
Lu: The paper sets the stage by showing how this game structure connects directly to how we generalize models when we see different data distributions.
Meng: So while the math is dense, the practical implication is that if you want a robust predictor for real-world problems, look into these invariant subsets because they handle uncertainty better.
Linus Kühne, Felix Schur, Jonas Peters
ETH Zurich
stat.ML, cs.AI, cs.LG, stat.ME
Submitted: 2026-05-16
Updated: 2026-11-04
Code: https://github.com/LinusKuehne/prediction-intervention-games
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 85/100
The gist: We consider the following two-player game: using observational data, the leader chooses a prediction function for a response variable Y from given covariates The single most informative sentence…
Key concepts
- Prediction-Intervention Games
- This formalizes a game where one player (the leader) makes a prediction function for an outcome, and another player (the follower) chooses an action (intervention). The leader tries to minimize the follower's potential gain, while the follower tries to maximize their own objective based on that prediction.
- Invariant Sets
- An invariant set is a collection of data points where the outcome Y is statistically independent of which specific intervention E occurs. This means if you know which subset S of covariates X is active, the conditional distribution of Y remains the same regardless of which specific intervention within that set happens.
- Stable Blanket (SB(Y))
- The stable blanket is a set that captures as much relevant information about Y as possible while excluding irrelevant covariates. It satisfies properties related to invariance and maximizes information retention, making it a superior predictor compared to simply using the causal parents.
Terminology
Summary
We consider the following two-player game: using observational data, the leader chooses a prediction function for a response variable Y from given covariates The single most informative sentence about the paper's subject and finding is that predictors based on the stable blanket are always better or as good as those based on the causal parents
Prediction-Intervention Games
The paper formalizes the problem as a causal version of a Stackelberg game, which it calls a prediction-intervention game, where the leader provides a prediction function f and the follower reacts by choosing an intervention e to maximize their objective The leader seeks to find argminf Rlead e(f) (f), while the follower selects e in argmaxe V foll e(f) (f) This setup is a special case of a Stackelberg game.
Invariant Sets and Predictors
The paper introduces the concept of an invariant subset S, defined as a subset where Y ⊥⊥ E XS under P, or equivalently, if Y XS is the same under all Pe, e ∈ E The stable blanket SB(Y) is defined as a set satisfying PA(Y) ⊆ SB(Y) ⊆ MB(Y), which captures as much information about Y as possible while excluding irrelevant covariates.
Optimality of the Stable Blanket
Theorem 3 states that for two common follower objectives, either V foll e(f) = Re(f) or V foll e(f) = Ee[α + β f(X) − ce(X)] for α, β ∈ R and a family of measurable functions on X that do not depend on f, then sup e∈E∗(fSB(Y)) Re(fSB(Y)) ≤ sup e∈E∗(fPA(Y)) Re(fPA(Y)) This proves that predictors based on the stable blanket are better or as good as those based on the causal parents under these conditions.
Worst-Case Distribution Generalization
The worst-case risk over the possible interventions upper-bounds the leader’s post-intervention risk and does not require knowledge of the follower’s objective The bound (4) shows that sup e∈E∗(f) Re(f) ≤ sup e∈E Re(f), which connects prediction-intervention games to distribution generalization. Theorem 5 proves that under Assumption A1 or A2, fSB(Y) solves Problem (5): sup e∈E Re(f) ≥ sup e∈E Re(fSB(Y)). p. 5 If either Asm. A1 or A2 holds, then for all measurable f : X → A, fSB(Y ) solves Problem (5):
sup
e∈E
Re(f) ≥ sup
e∈E
Re(fSB(Y )).Where the paper says it
Learning Subset-Based Predictors
The paper proposes learning methods where the leader searches for subsets S ⊆ d that are both invariant on Etr and predictive of Y, leading to the IMP (invariant most-predictive) approach This involves estimating a predictor fS(x) = E[Y XS = xS] trained on data pooled across training environments. Stabilized regression (SR) and stabilized classification (SC) methods are proposed to address the impact of errors in the invariance test. p. 6 Stabilized regression and classification. Since IMP is based on a single subset, errors in the
invariance test may have a large impact on IMP’s performance.Where the paper says it
Experimental Validation
The theoretical results are validated on synthetic data, showing that SB(Y) improves over PA(Y) in both linear-Gaussian and nonlinear SCMs Experiments on real data from the Causal Chambers light tunnel also confirm the findings, demonstrating that the stable blanket predictor is robust to inaccurate causal graph assumptions.
Implementation Details
The implementation details for both the leader's predictor and the follower's intervention mechanism are provided in sections E.4 and F.5, detailing how MLPs are used for estimation and optimization under specific objectives. p. 0 To avoid
severe performance loss, the leader may base their prediction on the causal parents of Y , or more generally on an invariant subset of covariates. We prove, for
two common classes of follower objectives, that predictors based on the stable
blanket, a specific invariant subset, are always better or as good as those based on
the causal parents. This is consistent with the
hidden-confounding hypothesis that the graph is incorrect. The SC predictor ˆf
SC, which is purely
based on data, selects {RGB, ir_2, vis_2} (the stable blanket under the hidden-confounder graph)
as the unique subset passing both the invariance and predictiveness filters, and therefore matches
6Specifically, ir_i ⊥⊥ ir_j | RGB and vis_i ⊥⊥ vis_j | RGB for i ̸= j ∈ {1, 2, 3} are rejected in the majority of
training environments at level 0.Where the paper says it
Where the paper says it
Proofs of Key Lemmas
Several lemmas are provided to support the main theorems, such as Lemma 13 which generalizes the orthogonality of conditional expectation to regression and binary classification with loss functions strictly consistent for the mean functional Proofs for Lemmata 2, 4, and 14 establish that SB(Y) is invariant by the global Markov property. The proof of Prop. p. 5 , Rojas-Carulla et al. 4] (or, more precisely, Prop. , Rojas-Carulla et al. 4] (or, more precisely, Prop. , Rojas-Carulla et al. 4] (or, more precisely, Prop. , Rojas-Carulla et al. 4] (or, more precisely, Prop. , Rojas-Carulla et al. 4] (or, more precisely, Prop. , Rojas-Carulla et al. 4] (or, more precisely, Prop. , Rojas-Carulla et al. 4] (or, more precisely, Prop. , Rojas-Carulla et al. 4] (or, more precisely, Prop. , Rojas-Carulla et al. 4] (or, more precisely, Prop. , Rojas-Carulla et al. 4] (or, more precisely, Prop. , Rojas-Carulla et al. 4] (or, more precisely, Prop. , Rojas-Carulla et al. 4] (or, more precisely, Prop. , Rojas-Carulla et al. 4] (or, more precisely, Prop. , Rojas-Carulla et al. 4] (or, more precisely, Prop. , Rojas-Carulla et al. 4] (or, more precisely, Prop. , Rojas-Carulla et al. 4] (or, more precisely, Prop. , Rojas-Carulla et al. 4] (or, more precisely, Prop. , Rojas-Carulla et al. 4] (or, more precisely, Prop. 5 is a function-level statement: under Asm. A1 or A2, no measurable predictor improves on
fSB(Y )
in worst-case risk. § 4 asks whether fSB(Y )
is worst-case optimal among all measurable predictors and connects this
question to standard distribution generalization. 4 Distribution generalization and an upper bound
§ 3 shows that no σ(XDE(Y )C
forb
)-measurable predictor improves on fSB(Y )
in any environment. We formulate our results using the former notation. 1 Optimality of the stable blanket
By Lem. We formulate our results using the former notation. 1 Optimality of the stable blanket
By Lem. We formulate our results using the former notation. 1 Optimality of the stable blanket
By Lem. We formulate our results using the former notation. 1 Optimality of the stable blanket
By Lem. We formulate our results using the former notation. 1 Optimality of the stable blanket
By Lem. We formulate our results using the former notation. 1 Optimality of the stable blanket
By Lem. Assume that for all j ∈ CHG0
(E), Aj ⊆ DEG0
(Y )
C
forb \ {Y, E}, and that G is acyclic. Then DEG(Y )forb = DEG0
(Y )forb and SBG(Y ) = SBG0
(Y ). (3)
Since E ⊥G Y | SB(Y ) holds, SB(Y ) is invariant by the global Markov property. In previous work, optimality of the stable blanket has been discussed only as an average optimality
among invariant sets [41, Thm. (3)
Since E ⊥G Y | SB(Y ) holds, SB(Y ) is invariant by the global Markov property. In previous work, optimality of the stable blanket has been discussed only as an average optimality
among invariant sets [41, Thm. (3)
Since E ⊥G Y | SB(Y ) holds, SB(Y ) is invariant by the global Markov property. In previous work, optimality of the stable blanket has been discussed only as an average optimality
among invariant sets [41, Thm. (3)
Since E ⊥G Y | SB(Y ) holds, SB(Y ) is invariant by the global Markov property. In previous work, optimality of the stable blanket has been discussed only as an average optimality
among invariant sets [41, Thm. (3)
Since E ⊥G Y | SB(Y ) holds, SB(Y ) is invariant by the global Markov property. In previous work, optimality of the stable blanket has been discussed only as an average optimality
among invariant sets [41, Thm. (3)
Since E ⊥G Y | SB(Y ) holds, SB(Y ) is invariant by the global Markov property. In previous work, optimality of the stable blanket has been discussed only as an average optimality
among invariant sets [41, Thm. (3)
Since E ⊥G Y | SB(Y ) holds, SB(Y ) is invariant by the global Markov property. In previous work, optimality of the stable blanket has been discussed only as an average optimality
among invariant sets [41, Thm. (3)
Since E ⊥G Y | SB(Y ) holds, SB(Y ) is invariant by the global Markov property. In previous work, optimality of the stable blanket has been discussed only as an average optimality
among invariant sets [41, Thm. (3)
Since E ⊥G Y | SB(Y ) holds, SB(Y ) is invariant by the global Markov property. In previous work, optimality of the stable blanket has been discussed only as an average optimality
among invariant sets [41, Thm. (3)
Since E ⊥G Y | SB(Y ) holds, SB(Y ) is invariant by the global Markov property. In previous work, optimality of the stable blanket has been discussed only as an average optimality
among invariant sets [41, Thm. (3)
Since E ⊥G Y | SB(Y ) holds, SB(Y ) is invariant by the global Markov property. In previous work, optimality of the stable blanket has been discussed only as an average optimality
among invariant sets [41, Thm. (3)
Since E ⊥G Y | SB(Y ) holds, SB(Y ) is invariant by the global Markov property. In previous work, optimality of the stable blanket has been discussed only as an average optimality
among invariant sets [41, Thm. (3)
Since E ⊥G Y | SB(Y ) holds, SB(Y ) is invariant by the global Markov property. In previous work, optimality of the stable blanket has been discussed only as an average optimality
among invariant sets [41, Thm. (3)
Since E ⊥G Y | SB(Y ) holds, SB(Y ) is invariant by the global Markov property. In previous work, optimality of the stable blanket has been discussed only as an average optimality
among invariant sets [41, Thm. (3)
Since E ⊥G Y | SB(Y ) holds, SB(Y ) is invariant by the global Markov property. In previous work, optimality of the stable blanket has been discussed only as an average optimality
among invariant sets [41, Thm. (3)
Since E ⊥G Y | SB(Y ) holds, SB(Y ) is invariant by the global Markov property. In previous work, optimality of the stable blanket has been discussed only as an average optimality
among invariant sets [41, Thm. (3)
Since E ⊥G Y | SB(Y ) holds, SB(Y ) is invariant by the global Markov property. In previous work, optimality of the stable blanket has been discussed only as an average optimality
among invariant sets [41, Thm. Would they need to anticipate which edges
strategic followers add once f is deployed? The next lemma shows that this is not necessary: under
weak assumptions on the Aj ’s the stable blanket is the same in G0 and G, so SB(Y ) can be read off
from the training-stage DAG G0 alone. Proofs of all results are in § D. 2 implies a subset-level formulation, recovering [41, Thm. For all f, we have
sup
e∈E∗(f)
Re(f) ≤ sup
e∈E
Re(f). (4)
The bound is attained when the follower selects environments to maximize V
foll
e
(f) = Re(f), and
can be loose when the objectives are different. choose deterministic values for CH(E), taking into account the observation XO = xO. For simplicity, let us assume that for all j ∈ CH(E), we have PA(j) ⊆ O ∪ CH(E), so the follower has
access to all parents of j when setting j’s value. (3)
Since E ⊥G Y | SB(Y ) holds, SB(Y ) is invariant by the global Markov property. In previous work, optimality of the stable blanket has been discussed only as an average optimality
among invariant sets [41, Thm. (3)
Since E ⊥G Y | SB(Y ) holds, SB(Y ) is invariant by the global Markov property. In previous work, optimality of the stable blanket has been discussed only as an average optimality
among invariant sets [41, Thm. (3)
Since E ⊥G Y | SB(Y ) holds, SB(Y ) is invariant by the global Markov property. In previous work, optimality of the stable blanket has been discussed only as an average optimality
among invariant sets [41, Thm. (3)
Since E ⊥G Y | SB(Y ) holds, SB(Y ) is invariant by the global Markov property. In previous work, optimality of the stable blanket has been discussed only as an average optimality
among invariant sets [41, Thm. (3)
Since E ⊥G Y | SB(Y ) holds, SB(Y ) is invariant by the global Markov property. In previous work, optimality of the stable blanket has been discussed only as an average optimality
among invariant sets [41, Thm. 3The answer does not follow from existing work: e. , Rojas-Carulla et al. 3The answer does not follow from existing work: e. , Rojas-Carulla et al. 3The answer does not follow from existing work: e. , Rojas-Carulla et al. 3The answer does not follow from existing work: e. , Rojas-Carulla et al. 3The answer does not follow from existing work: e. , Rojas-Carulla et al. § 4 asks whether fSB(Y )
is worst-case optimal among all measurable predictors and connects this
question to standard distribution generalization. 4 Distribution generalization and an upper bound
§ 3 shows that no σ(XDE(Y )C
forb
)-measurable predictor improves on fSB(Y )
in any environment. § 4 asks whether fSB(Y )
is worst-case optimal among all measurable predictors and connects this
question to standard distribution generalization. 4 Distribution generalization and an upper bound
§ 3 shows that no σ(XDE(Y )C
forb
)-measurable predictor improves on fSB(Y )
in any environment. § 4 asks whether fSB(Y )
is worst-case optimal among all measurable predictors and connects this
question to standard distribution generalization. 4 Distribution generalization and an upper bound
§ 3 shows that no σ(XDE(Y )C
forb
)-measurable predictor improves on fSB(Y )
in any environment. § 4 asks whether fSB(Y )
is worst-case optimal among all measurable predictors and connects this
question to standard distribution generalization. 4 Distribution generalization and an upper bound
§ 3 shows that no σ(XDE(Y )C
forb
)-measurable predictor improves on fSB(Y )
in any environment. § 4 asks whether fSB(Y )
is worst-case optimal among all measurable predictors and connects this
question to standard distribution generalization. 4 Distribution generalization and an upper bound
§ 3 shows that no σ(XDE(Y )C
forb
)-measurable predictor improves on fSB(Y )
in any environment. § 4 asks whether fSB(Y )
is worst-case optimal among all measurable predictors and connects this
question to standard distribution generalization. 4 Distribution generalization and an upper bound
§ 3 shows that no σ(XDE(Y )C
forb
)-measurable predictor improves on fSB(Y )
in any environment. § 4 asks whether fSB(Y )
is worst-case optimal among all measurable predictors and connects this
question to standard distribution generalization. 4 Distribution generalization and an upper bound
§ 3 shows that no σ(XDE(Y )C
forb
)-measurable predictor improves on fSB(Y )
in any environment. § 4 asks whether fSB(Y )
is worst-case optimal among all measurable predictors and connects this
question to standard distribution generalization. 4 Distribution generalization and an upper bound
§ 3 shows that no σ(XDE(Y )C
forb
)-measurable predictor improves on fSB(Y )
in any environment. § 4 asks whether fSB(Y )
is worst-case optimal among all measurable predictors and connects this
question to standard distribution generalization. 4 Distribution generalization and an upper bound
§ 3 shows that no σ(XDE(Y )C
forb
)-measurable predictor improves on fSB(Y )
in any environment. § 4 asks whether fSB(Y )
is worst-case optimal among all measurable predictors and connects this
question to standard distribution generalization. 4 Distribution generalization and an upper bound
§ 3 shows that no σ(XDE(Y )C
forb
)-measurable predictor improves on fSB(Y )
in any environment. Suppose V
foll satisfies either (i) V
foll
e
(f) = Re(f), or (ii)
V
foll
e
(f) = Ee[ α + β f(X) − ce(X) ] for α, β ∈ R and a family {ce}e∈E of measurable functions
on X that do not depend on f. Then,
sup
e∈E∗(fSB(Y ))
Re(fSB(Y )) ≤ sup
e∈E∗(fPA(Y ))
Re(fPA(Y )). Suppose V
foll satisfies either (i) V
foll
e
(f) = Re(f), or (ii)
V
foll
e
(f) = Ee[ α + β f(X) − ce(X) ] for α, β ∈ R and a family {ce}e∈E of measurable functions
on X that do not depend on f. Then,
sup
e∈E∗(fSB(Y ))
Re(fSB(Y )) ≤ sup
e∈E∗(fPA(Y ))
Re(fPA(Y )). Suppose V
foll satisfies either (i) V
foll
e
(f) = Re(f), or (ii)
V
foll
e
(f) = Ee[ α + β f(X) − ce(X) ] for α, β ∈ R and a family {ce}e∈E of measurable functions
on X that do not depend on f. Then,
sup
e∈E∗(fSB(Y ))
Re(fSB(Y )) ≤ sup
e∈E∗(fPA(Y ))
Re(fPA(Y )). Suppose V
foll satisfies either (i) V
foll
e
(f) = Re(f), or (ii)
V
foll
e
(f) = Ee[ α + β f(X) − ce(X) ] for α, β ∈ R and a family {ce}e∈E of measurable functions
on X that do not depend on f. Then,
sup
e∈E∗(fSB(Y ))
Re(fSB(Y )) ≤ sup
e∈E∗(fPA(Y ))
Re(fPA(Y )). Suppose V
foll satisfies either (i) V
foll
e
(f) = Re(f), or (ii)
V
foll
e
(f) = Ee[ α + β f(X) − ce(X) ] for α, β ∈ R and a family {ce}e∈E of measurable functions
on X that do not depend on f. Then,
sup
e∈E∗(fSB(Y ))
Re(fSB(Y )) ≤ sup
e∈E∗(fPA(Y ))
Re(fPA(Y )). Would they need to anticipate which edges
strategic followers add once f is deployed? The next lemma shows that this is not necessary: under
weak assumptions on the Aj ’s the stable blanket is the same in G0 and G, so SB(Y ) can be read off
from the training-stage DAG G0 alone. Would they need to anticipate which edges
strategic followers add once f is deployed? The next lemma shows that this is not necessary: under
weak assumptions on the Aj ’s the stable blanket is the same in G0 and G, so SB(Y ) can be read off
from the training-stage DAG G0 alone. Would they need to anticipate which edges
strategic followers add once f is deployed? The next lemma shows that this is not necessary: under
weak assumptions on the Aj ’s the stable blanket is the same in G0 and G, so SB(Y ) can be read off
from the training-stage DAG G0 alone. Would they need to anticipate which edges
strategic followers add once f is deployed? The next lemma shows that this is not necessary: under
weak assumptions on the Aj ’s the stable blanket is the same in G0 and G, so SB(Y ) can be read off
from the training-stage DAG G0 alone. Would they need to anticipate which edges
strategic followers add once f is deployed? The next lemma shows that this is not necessary: under
weak assumptions on the Aj ’s the stable blanket is the same in G0 and G, so SB(Y ) can be read off
from the training-stage DAG G0 alone. In the special additive-noise case Y =
g(XPA(Y )) +ε with ε ⊥⊥ XPA(Y ) and E[ε] = 0, the follower’s interventions are inconsequential for
parent-based prediction under the squared loss: for all e ∈ E, Re(fPA(Y )) = Var(ε). Without this
additive-noise structure, the risk may still vary across environments through the marginal distribution
of XPA(Y )
, but the variation is only due to covariate shift under a fixed conditional Y | XPA(Y )
. In the special additive-noise case Y =
g(XPA(Y )) +ε with ε ⊥⊥ XPA(Y ) and E[ε] = 0, the follower’s interventions are inconsequential for
parent-based prediction under the squared loss: for all e ∈ E, Re(fPA(Y )) = Var(ε). Without this
additive-noise structure, the risk may still vary across environments through the marginal distribution
of XPA(Y )
, but the variation is only due to covariate shift under a fixed conditional Y | XPA(Y )
. In the special additive-noise case Y =
g(XPA(Y )) +ε with ε ⊥⊥ XPA(Y ) and E[ε] = 0, the follower’s interventions are inconsequential for
parent-based prediction under the squared loss: for all e ∈ E, Re(fPA(Y )) = Var(ε). Without this
additive-noise structure, the risk may still vary across environments through the marginal distribution
of XPA(Y )
, but the variation is only due to covariate shift under a fixed conditional Y | XPA(Y )
. In the special additive-noise case Y =
g(XPA(Y )) +ε with ε ⊥⊥ XPA(Y ) and E[ε] = 0, the follower’s interventions are inconsequential for
parent-based prediction under the squared loss: for all e ∈ E, Re(fPA(Y )) = Var(ε). Without this
additive-noise structure, the risk may still vary across environments through the marginal distribution
of XPA(Y )
, but the variation is only due to covariate shift under a fixed conditional Y | XPA(Y )
. (4)
The bound is attained when the follower selects environments to maximize V
foll
e
(f) = Re(f), and
can be loose when the objectives are different. Minimizing the right-hand side is a distribution
generalization (DG) problem. (4)
The bound is attained when the follower selects environments to maximize V
foll
e
(f) = Re(f), and
can be loose when the objectives are different. Minimizing the right-hand side is a distribution
generalization (DG) problem. (4)
The bound is attained when the follower selects environments to maximize V
foll
e
(f) = Re(f), and
can be loose when the objectives are different. Minimizing the right-hand side is a distribution
generalization (DG) problem. (4)
The bound is attained when the follower selects environments to maximize V
foll
e
(f) = Re(f), and
can be loose when the objectives are different. Minimizing the right-hand side is a distribution
generalization (DG) problem.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Conclusion
The paper concludes that the subset view is arguably more interpretable and transparent than function-based invariance criteria like IRM. The subset view combines feature selection with standard supervised learning, simplifying auditing and making individual predictions easier to explain. The paper also notes that the results rely on the assumption of strong interventions, and finite-sample analysis is left for future work. p. 9 Finally, in
practice, the leader uses a finite-sample estimate ˆfS rather than the population predictor fS, and the
follower can shift covariates to high-error regions (§ E. 3); a finite-sample analysis is left for future
work. This is consistent with the
hidden-confounding hypothesis that the graph is incorrect. The SC predictor ˆf
SC, which is purely
based on data, selects {RGB, ir_2, vis_2} (the stable blanket under the hidden-confounder graph)
as the unique subset passing both the invariance and predictiveness filters, and therefore matches
6Specifically, ir_i ⊥⊥ ir_j | RGB and vis_i ⊥⊥ vis_j | RGB for i ̸= j ∈ {1, 2, 3} are rejected in the majority of
training environments at level 0. We further upper bound the leader’s post-intervention risk by
a worst-case risk over allowed interventions and strengthen existing distribution
generalization results to analyze this bound: we give sufficient conditions under
which stable-blanket predictors are worst-case optimal, and show by examples
that these conditions cannot in general be dropped. Finally, we discuss practical
strategies for settings with known and unknown graph, and test them on simulated
and real-world data.Where the paper says it
Where the paper says it
Where the paper says it
Appendix Contents
The appendix contains extended related work, additional theoretical results, stabilized regression and classification details, proofs of the theorems, and additional details for the synthetic and Causal Chambers experiments. The appendix includes detailed descriptions of invariance tests like IRD and ITP. p. 7 first to port this idea to classification by replacing the base regressors, the prediction score (crossentropy instead of MSE), and the invariance tests with their classification counterparts, yielding
stabilized classification (SC). Details on SC are given in § C. 3); a finite-sample analysis is left for future
work. Acknowledgments and Disclosure of Funding
We thank Niklas Pfister and Sorawit Saengkyongam for helpful discussions, and Juan Gamella for
the help with the Causal Chambers data generation.Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improved AI systems can select predictors based on invariant subsets, such as stable blankets, which are proven to be
always better or as good as those based on the causal parents
for two common classes of follower objectives. This allows the leader to design robust prediction functions that are less susceptible to manipulation by followers who intervene on covariates. -
AI systems can adopt learning methods like Stabilized Regression and Classification, which involve a two-stage procedure to select subsets of covariates that are both invariant on training data and predictive of Y, leading to a final ensemble predictor. This approach is robust when the causal graph is unknown, as it relies only on training data from several environments.
-
The system can utilize an upper bound for the leader’s post-intervention risk, which
holds irrespective of V foll
andallows for easier optimization than the game objective.
This provides a concrete worst-case risk that is independent of the follower’s unknown objective, connecting prediction-intervention games to distribution generalization. -
The system can perform inference in settings where the causal graph is unknown by learning methods that search for subsets S that are both invariant on training data and predictive of Y, using invariance scores like sinv(S) and prediction scores spred(S).
Abstract
We consider the following two-player game: using observational data, the leader chooses a prediction function for a response variable Y from given covariates. The follower then reacts with an intervention on some covariates in the underlying structural causal model to maximize their own objective. The leader knows the intervention targets, but may have limited knowledge of the follower's objective. We call this setup a prediction-intervention game, a special case of a Stackelberg game. Finding an optimal strategy for the leader is generally difficult. To avoid severe performance loss, the leader may base their prediction on the causal parents of Y, or more generally on an invariant subset of covariates. We prove, for two common classes of follower objectives, that predictors based on the stable blanket, an invariant subset, are at least as good as those based on the causal parents. We further upper bound the leader's post-intervention risk by a worst-case risk over allowed interventions and strengthen existing distribution generalization results to analyze this bound: we give sufficient conditions under which stable-blanket predictors are worst-case optimal, and show by examples that these conditions cannot in general be dropped. Finally, we discuss practical strategies for settings with known and unknown graph, and test them on simulated and real-world data.
Sources
- Invariant Risk Minimization
- Fair lending needs explainable models for responsible recommendation
- Linear Classifiers that Encourage Constructive Adaptation
- Maximum Risk Minimization with Random Forests
- Incentivizing High-Quality Content in Online Recommender Systems
- Identifying the Causes of Pyrocumulonimbus (PyroCb)
- Stable Blanket with Hidden Variables and Cycles
- Discovering Optimal Scoring Mechanisms in Causal Strategic Prediction
Related papers
- Uniform-in-Time Weak Propagation-of-Chaos in Shallow Neural Networks
- Diffusion Models Observe Only Gradients: A Geometric Perspective on Score Matching Errors
- Quantifying Error Propagation and Model Collapse in Diffusion Models
- Behavior of prediction performance metrics with rare events
- Statistical Inference and Learning for Shapley Additive Explanations (SHAP)
- Control Variate Design for Score-Function Gradient Estimators in Importance-Weighted Variational Inference