Intervention Generalization: A View from Factor Graph Models

One of the goals of causal inference is to generalize from past experiments and observational data to novel conditions. While it is in principle possible to eventually learn a mapping from a novel experimental condition to an outcome of interest, provided a sufficient variety of experiments is available in the training data, coping with a large combinatorial space of possible interventions is hard. Under a typical sparse experimental design, this mapping is ill-posed without relying on heavy regularization or prior distributions. Such assumptions may or may not be reliable, and can be hard to defend or test. In this paper, we take a close look at how to warrant a leap from past experiments to novel conditions based on minimal assumptions about the factorization of the distribution of the manipulated system, communicated in the well-understood language of factor graph models. A postulated $\textit{interventional factor model}$ (IFM) may not always be informative, but it conveniently abstracts away a need for explicitly modeling unmeasured confounding and feedback mechanisms, leading to directly testable claims. Given an IFM and datasets from a collection of experimental regimes, we derive conditions for identifiability of the expected outcomes of new regimes never observed in these training data. We implement our framework using several efficient algorithms, and apply them on a range of semi-synthetic experiments.

Paper

Similar papers

Peer review

Reviewer 2Na96/10 · confidence 3/52023-07-03

Summary

This paper studies the problem of generalization in causal inference. In particular, it extends the factor graph to interventional factor graph (IFM). It shows when can such model be identified as well as proving practical algorithm for learning. The setting assumes knowing the factorization, this type of structural knowledge has its advantages and limitations as well.

Strengths

The paper is well motivated with good theoretical results and empirical experiments.

Weaknesses

1. This paper does not provide too much real-world example of the interventional factor model. The simulation is also semi-synthetic. I am a little concerned about the applicability of such model. 2. There seems to be a disconnect between the identifiability results and learning algorithms (see the questions section).

Questions

1. Because the term junction tree is explicitly mentioned in theorem 3.1, it would be good to have a brief definition of it? 2. Can you elaborate a bit more on the algebraic formulation of this problem? This section is a bit harder to digest. 3. I am a little confused about the usefulness of theorem 3.1 and 3.2. As authors mentioned in the first paragraph of Section 4, they do not use it to product of density ratios but rather use deep energy-based model or IPW. Are the identifiability results in 3.1 and 3.2 necessary for these methods to work?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

Not applicable

Reviewer EjuP5/10 · confidence 4/52023-07-07

Summary

This work proposes the use of factor models as a graphical causal model to generalize from past experiments. The authors introduce factor models, describe their relative merit then describe how a factorization can be derived for a given intervention, and give approaches for estimation using deep energy based modeling and weighting approaches. Experimental results are shown which provide a comparison of the performance of the proposed approach to other approaches (DAGs and black box estimation)

Strengths

* This is a very interesting idea, and I generally agree with the authors' central claims around the opportunities provided by using factor models for causal inference. * The addition of conformal inference for uncertainty intervals here is a nice, and elegant addition to the paper. * The authors do a nice job of providing a thorough empirical evaluation of the proposed approach

Weaknesses

One of the weaknesses of factor models is that it is more difficult to perform inference than in DAGs where there is a simple factorization that can be exploited. Factor models are also less immediately interpretable than DAG models. While this isn't necessary a problem in itself, it would be useful if there was a more plain discussion about the tradeoffs involved in this representation. I also found the presentation to be a little difficult to follow. There are a number of missing discussions that would be useful to contextualize the proposed approach in the broader literature (e.g., ADMGs, segregated graphs, gated factor graphs). It's also a little unclear to me whether there is a sound and complete identification algorithm here.

Questions

* Is the set of identifiable estimands comparable to other graphical causal models like ADMGs? * For these estimation approaches (and more generally) is it possible to provide a sense of convergence/consistency of the causal parameters? * Given the experimental results where there does not seem to be a clearly preferable approach in all settings, how should someone decide when it is appropriate/necessary to employ an IFM? * Can you provide a discussion of the current proposal with "Causality with Gates" by Winn (AISTATS 2012). I can see that there are significant differences between these two texts, but given that they are both concerned with the use of factor graphs for causal inference I think it should be discussed.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

Yes.

Authorsrebuttal2023-08-18

Can we help with more comments?

Thank you for all the feedback so far. Please do let us know if there is anything else for us to address in our rebuttal. All the best, the authors

Reviewer t5cb6/10 · confidence 2/52023-07-07

Summary

The authors present the interventional factor model, a more general formalization used to predict the effect of treatment on an outcome in unseen regimes. It is more general than existing formalizations since it does not assume causal graphs to be directed acyclic graphs (DAGs). The authors show in this new formalization what are the conditions to ensure identifiability of treatment effects. Finally, the authors propose several methods that can estimate treatment effects and show their effectiveness with semi-synthetic experiments.

Strengths

The article is well written. It proposes a really general formalization using factor graphs that is original. It encompasses identification results, several algorithms, and experiments.

Weaknesses

The contribution is limited in the sense that identification results for DAGs are already well established (do-calculus and $\sigma$-calculus). This work address the more general case where the graph is not necessarily a DAG. It supposes that the graph is known, but this assumption is strong in practice since it is challenging to know these kinds of general graphs both from the expert knowledge and structure learning perspective (more than DAGs). Real-world applications could surely help motivate the use of this formalization. The present semi-synthetic experiments are interesting, however the simple black-box baseline method is overall performing really well, undermining the use of the more involved proposed methods.

Questions

Minor typos or style suggestions: - line 39: "we submit" => "we argue" or "we claim" - line 46: use a colon instead of a full stop - line 59: why use the aleph symbol, could another common letter be used instead? - line 204: "Equ." => "Eq.". It is more common and "Eq." is used for all the other references to equations. - line 350: citation 2 is repeated

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

Yes the limitation have been addressed, and the societal impact is not really applicable.

Reviewer 4dmo7/10 · confidence 3/52023-07-07

Summary

This paper consider how to use data from past interventions to allow it to generalize to new unseen interventions. This is an important practical problem to consider, as running additional experiments is often costly/infeasible. In order to tackle this problem, the authors consider a graphical models approach, specifically they consider an interventional factor graph model (IFM). Using posit an IFM factorization of the the density $p(x; \sigma)$. Then they provide the sufficient conditions for identification, and provide a message passing algorithm to do so. Next, they discuss multiple approaches to estimate the density based on ML models, as as well via IPW methods. They also consider a covariate shift regression approach. Further, they provide a conformal inference approach to establish coverage. Lastly, they perform a number of experiments to establish the empirical efficacy of their method.

Strengths

The paper tackles a very important problem, how can we use past interventions to generalize to new interventions. They propose an innovative IFM model and message passing algorithm to establish identification for this problem. Viewing this problem under this lens is an interesting one, and potentially of practical use. Further, I appreciate the authors providing examples so that its easier to understand their identification argument. I also appreciated the authors providing multiple methods for estimating the density discussed earlier. I believe providing multiple approaches is often of great practical use since no one algorithm often works in all scenarios. In terms of empirical evaluation, I think its interesting that the authors used semi-synthetic data. I believe this is good practice, and should be followed more often.

Weaknesses

Presentation: Presentation of both the regression and coverage algorithm is confusing, and seems to require a lot of additional knowledge on behalf of the reader. I do not fully understand how these algorithms proceed. For example, the deep-energy based models, lines 237-239 are very unclear. Making it clear what exactly is being fit would be very useful. More generally, being clear and rigorous regarding these things will go a long way in making the paper clearer. Empirical Evaluation of Coverage: I did not see any simulations to this effect. Empirical Evaluation: It is not clear to me what exactly X is in these datasets. Once again, being clear and rigorous about these details will enhance understanding, and give the reader a chance to appreciate the empirical evaluation. Further, even after reading the appendix, I do not understand how the outcomes were generated. Could the authors please clarify empirical details in the rebuttal? Comparison to related work: The comparison to [2] is incorrect. The authors claim that a series of works including that of [2] requite data to be collected for all regimes in $\Sigma_{\text{test}}$. This is not the case. For example, the experimental design section of [2] shows that this isn't the case.

Questions

I have made some suggestions in the weaknesses section. I list some other questions here. Deep-energy based models: Does fitting the parameter vector $\theta_{k, \sigma_{F_{k}}}$ require knowledge of the set of variables in $F_k$? I am confused by this, and if it does require, how do we determine these variables in practice.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

Yes, they have.

Reviewer ybr15/10 · confidence 3/52023-07-26

Summary

This paper maps from available observational and experimental datasets to unseen interventional distributions given the factorization of the joint distribution of the intervened system. They utilize an interventional factor model equipped with factor graphs to provide necessary and sufficient conditions for causal effect identifiability. Finally, they provide some practical algorithms to estimate the final outcomes.

Strengths

The authors provided good examples of different concepts in the main paper and in the appendix. It's a different approach from approaches that use causal graphs but is quite interesting. The paper has a nice flow in the writing and is easy to read. The explanations and detailed examples provided in the appendix are worth appreciating, and they are quite helpful for readers.

Weaknesses

Main weakness: * Some concepts such as junction tree, hypervertex, and message passing algorithms should have been defined with short examples in the main paper since they are used in the main paper theorems. * Multiple approaches have been described in sections 3 and 4 and but most of them were not explained with enough details and intuition. * The main contribution seems a little unclear. The authors first discussed two approaches in section 3. Later in section 4, they mentioned that in practice using deep energy-based models works better. They suggested employing a differentiable black box to learn parameters for each factor and estimate E[Y| x]. Using a black box is not completely novel. It appears that the approaches in section 3 are not very useful and thus the authors are proposing three more methods that work better in practice. I would request the authors to clarify the mentioned issues. * In the experiment section, the authors completely ignore the approaches they discussed in Theorem 3.1 or Theorem 3.2. They considered the deep learning approach as the best version and compared the benchmarks with that. Minor comments: * Some concepts in sections 1 and 2 are used without proper definitions and examples. Readers would need to know those background knowledge beforehand to go with the flow of the paper. * Line 119: Unmeasured confounders are used without any definitions or proper examples. * Line 207: Theorem 3.2 seems less intuitive. A proof sketch in the main paper would be appreciated. * Line 227: The approach “Deep energy-based models and direct regression” should be described in more detail since this approach worked better than other approaches. * Conformity scores are not defined in the main paper although has been used in theorem 4.1. I have read the author's rebuttal. The authors resolved some of my concerns. But I am not confident enough to increase the scores.

Questions

I would request the authors to provide explanations for the previously mentioned main weaknesses and answer the following questions: * For the example in line 74, why is the DAG $\sigma \rightarrow X$, $\sigma \rightarrow Y \leftarrow X$, not an option? * Are the authors refuting the utility of theorem 3.1 and 3.2 and adopting only the new approaches provided in section 4? * How is this paper dealing with cycles and confounders? The authors should mention that more explicitly.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

If the variables have more than two states or we have more variables, the number of interventional datasets will also be significantly high. In real-life applications, collecting these many interventional datasets are not feasible or possible. The authors should discuss this challenge in detail.

Reviewer T46G6/10 · confidence 4/52023-07-27

Summary

The paper introduce Interventional Factor Models (IFM), a graphical model that encodes assumptions about a data-generating process and the interventions that can be performed on top of it. The model explicitly includes intervention variables to impose a factorization over the distributions generated by the model under any interventional regime. Being a variable Y a function of the other observable variables X (or at least independent of the interventional regimes given X), the problem setting aims to learn E[Y] under a particular interventional regime of interest, given a subset of all possible interventional distributions and a corresponding IFM. For this task, the paper presents identification criteria and learning/estimation algorithms to find the parameter of interest. The authors present the result of experiments on semi-synthetic data to evaluate the performance of their approach. I have some questions related to the scope of the results in the paper. I leave them in the comments section and would like to hear from the authors. I'm willing to update my scores based on the response.

Strengths

- The paper presents a novel graphical model that can be used to generalize known experimental settings to new unseen settings of interest. - Relevant setup, definitions, and results needed to understand the result are included. - The writing is clear enough to understand the setting and main contributions of the paper. - The paper presents sound results for the identification of causal effects and algorithms to estimate it, based on seen experimental settings and an IFM.

Weaknesses

- The introduction and contributions claim that Thm. 3.2 is necessary, in general. However, the statement talks about PR-transformations in particular. It is not clear to me if it is proven that there are no other identification routes than PR-transformations. - It is mentioned in the paper that the IFM graph could be elicited from domain experts. This is something that sounds natural to me in the case of DAG, but not so intuitive in this case. I believe the reader could benefit from a more explicit example where variables have associated meaning and one could make sense of the assumptions.

Questions

- When Thm. 3.2 says no, is it the case that there is no way to identify the causal effect of interest from the given model and distributions, even if it is not in the form of a PR-transformation? - If not, does it mean that Thm. 3.2. is necessary only under PR-transformations?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

I believe the discussion on limitations and societal impacts is sufficient.

Reviewer t5cb2023-08-11

I extend my gratitude to the reviewers for their comprehensive response. I tend to think that this paper should be accepted.

Reviewer 4dmo2023-08-11

Response to Rebuttal

Thank you for the comments, and clarifications. I still believe that a score of 7 is a fair score of this paper, and I will keep it as such.

Reviewer ybr12023-08-16

Thanks to the authors for their clarifications. I would like to hear the authors' opinions and explanations about the limitations I have mentioned for this paper such as the high number of interventional datasets requirement.

Authorsrebuttal2023-08-17

Number of datasets

Thanks for the opportunity to follow-up on this (we ran out of space in the original box, and we thought it would be confusing to use a global response.) When the number of values per variable is particular high, we can parameterize a potential function as a smooth function of the values of each intervention variable. The identifiability results can show whether for a particular training grid of points we can identify (say) the product space over the combination of training intervention values, and uncertainty quantification methods can be used to indicate to which extent we have information to interpolate/smooth over test treatment levels points that lie in between the training levels. (To illustrate this with an analogy to the DAG case, imagine that we had each conditional distribution for a given random variable not as set of independent regressions - one for each combination of treatment parents values - but instead as a smooth mapping such as a Gaussian process with a real-valued encoding of the treatments as input) For a large number of variables, this is a generally hard problem overall regardless of method. However, the identification problem is with respect to a test set that does not need to span all possible combinations of interventions to be useful (e.g. what to focus on may be even lower-order interactions, such as in the pairwise DREAM analysis we use as illustration), with the junction tree approach showing a divide-and-conquer structure of subproblems which can be solved without providing solutions to larger joint set of variables. Many thanks again!

Reviewer T46G2023-08-17

Thanks for addressing my questions and comments. I still feel that the paper needs more concrete examples in terms of connecting with real-world systems. This is especially important when introducing a new family of graphical models that comes with a particular kind of implied assumptions. This gives the reader not only a better understanding of the model, but also a sense of how could those models be elicited in practice and the kind of systems they could be more useful. Together with the previous point, the concrete examples could help understanding the other limitation of the results, namely, needing an outcome Y that becomes independent of the treatment regimes given the other variables. The possible impact of the results critically depends on whether interesting systems fit this assumption, and if there are advantages in using IFM+results instead of other models. Having said that, I'm raising my score to a 6 after the authors response.

Reviewer EjuP2023-08-21

Thank you for your thoughtful response, and apologies for a delayed reply. I appreciate your framing on the generality of factor models versus DAGs, however it still isn't entirely clear to me when a practitioner should prefer to use the factor graph over existing frameworks, especially since identification is out of scope of this paper (contrast this to e.g., ADMGs which do admit identification). With that being said, I feel that the authors response does alleviate at least some of my concerns and I am upgrading my score to reflect this.

Authorsrebuttal2023-08-21

Thank you!

Thanks for the further consideration, we are mindful of real-world constraints and we appreciate the feedback at any stage! To be honest, we are not totally clear which missing identification results are being referred to, but no hurry. Any further detail you may be able to provide us at some point at the time decisions are released will be welcome and we will take them into full consideration. Many thanks again!

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC