Do Not Marginalize Mechanisms, Rather Consolidate!

Structural causal models (SCMs) are a powerful tool for understanding the complex causal relationships that underlie many real-world systems. As these systems grow in size, the number of variables and complexity of interactions between them does, too. Thus, becoming convoluted and difficult to analyze. This is particularly true in the context of machine learning and artificial intelligence, where an ever increasing amount of data demands for new methods to simplify and compress large scale SCM. While methods for marginalizing and abstracting SCM already exist today, they may destroy the causality of the marginalized model. To alleviate this, we introduce the concept of consolidating causal mechanisms to transform large-scale SCM while preserving consistent interventional behaviour. We show consolidation is a powerful method for simplifying SCM, discuss reduction of computational complexity and give a perspective on generalizing abilities of consolidated SCM.

Paper

References (30)

Scroll for more · 18 remaining

Similar papers

Peer review

Reviewer ZzBw7/10 · confidence 3/52023-07-04

Summary

The paper presents a framework for causal reasoning that supports the simplification of large structural causal models (SCMs). One key operation that supports this simplification is consolidation of a SCM such that only a subset of endogenous variables are explicitly modelled, but in such a way that all possible interventions are still supported. Another key operation is the partitioning of an SCM into sub-models, in such a way that endogenous variables from one sub-SCM act as exogenous variables in another. Potential reductions in complexity are then discussed by the use of these operations and associated constraints. For instance, replacing equations within SCMs with computationally simpler expressions, and dropping equations which do not relate to variables of interest. Throughout there is a focus on retaining information about the effects of interventions. The paper concludes with two examples: one on modelling tool wear on a milling machine and another, more complex example, exploring planning policies for a simple platformer game.

Strengths

* The paper very clearly presents its contribution and relates this well to existing work, justifying the relevance of the contribution to the field. * The paper builds its arguments formally, and with intuition, and presents meaningful relevant results and constraints. * The examples illustrate well the benefits of the proposed framework. * The contributions appear meaningful to me and likely to be of relevance to other researchers, particularly those concerned with large scale causal modelling.

Weaknesses

* At times the arguments are a little vague or the explanations incomplete. * Although mostly clear, the formal notation sometimes leaves a little to be desired. * There are a few places where the authors appear to make errors or omissions in their explanations. Issues with understanding: 1. In definition 1, the set of possible interventions is a little unclear. There appears to be 1 possible intervention, $I_i$ per endogenous variable, $X_i$. Or can there be multiple potential, but mutually exclusive, interventions per endogenous variable? What exactly is $I_i$? 2. On line 94 $f_i(\textbf{x}, x_0)$ is used to indicate the structural equation for variable $X_i$ under the intervention that $X_j$ is set to value $x_0$, but this seems a little under-defined to me. 3. On line 98, the authors state that $\mathcal{M}$ entails infinitely many intervened distributions, but this seems to conflict with earlier notation (see 1.). 4. Definition 2 could do with an explanation of $P^{\textbf{I}}_{\textbf{E}}$. I am assuming this means the distribution of target variables in $\textbf{E}$ under some intervention $\textbf{I}$ but I can't see this stated anywhere. 5. The notation $\rho(\textbf{U},\textbf{I})$ first introduced in Definition 2 does not refer to the subset of variables $\textbf{E}$ to which it refers. This would be good practice anyway, but becomes more problematic when partially consolidated SCMs are considered (Def 5) as the variables of interest $\textbf{E}$ are augmented with additional variables that act as exogenous variables in other sub-SCMs. 6. The caption for Figure 2 (right) states that the dotted line indicates explicit computation for $X_2$ but it isn't clear what is meant by this. 7. In section 3, in a number of places (lines 159, 165, 187), there appears to be repeated errors in the notation, e.g. $\textbf{V} \in \textbf{A}$. I think that $\textbf{V}$ is the complete set of endogenous elements while $\textbf{A}$ is a subset of $\textbf{V}$ 8. In lines 175-183, the partitioning of an SCM appears to require that exogenous variables $\textbf{U}_i$ of a sub SCM must be endogenous variables of another sub-SCM in the same partitioned SCM. But could they be truly exogenous variables of the whole system? Also, there appears to be notational irregularities in this paragraph relating to what is a sub-SCM and what is a partitioned SCM. 9. Things get a bit messy around definition 5 with respect to $\textbf{E}$ and $\textbf{E}'$. The distinction between these two sets of variables could be clearer. For instance, if I am considering partially consolideted SCM $\mathcal{M}_{\mathcal{A},\textbf{E}'}$ then how do I know what the set $\textbf{E}$ is that is used to define $\textbf{E}'$. 10. In section 4, there is a chain of inference on lines 234-235 that is difficult to follow (what is the scope of the universal and existential quantifiers?) and some entity $D$ appears without being defined. 11. I got a little lost in section 4.1. In particular, the discussion of conditional branching and stacking was a little vague. ## Post rebuttal After reading the rebuttal and individual responses to the above points, I am raising my recommendation to "accept".

Questions

My questions have mostly been articulated in the **Weaknesses** field.

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

2 fair

Presentation

3 good

Contribution

3 good

Limitations

The authors discuss limitations and the possible relaxations of these effectively.

Reviewer tWPL6/10 · confidence 3/52023-07-06

Summary

This work introduced a concept of consolidating causal mechanisms to transform large-scale Structural causal models (SCMs) while preserving consistent interventional behaviour. The author shows consolidation is powerful for simplifying SCM, disscuss the complexity and give a perspective on generalization.

Strengths

The author builds a solid framework of consolidating causal mechanisms. The expression of notations is adeque. The spirit of compression the causal equation is interesting and straight-forward.

Weaknesses

1. The author should spend more efforts to describe the compression of causal equations clearly, as this is the key contribution of this work. 2. The author did not propose a concrete algorithm to compress the causal equation.

Questions

1. Can author make a detail derivation on how to extract a consolidating causal mechanism from a toy SCM? 2. In what scale of SCMs can this method handle, and what is the complexity? Can author give some larger scale examples?

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

2 fair

Contribution

3 good

Limitations

see the weakness and question parts.

Reviewer CKKq6/10 · confidence 3/52023-07-07

Summary

An operation of consolidation on SCMs is defined and its merits laid out in numerous examples. The operation amalgamates variables while preserving aspects of the causal structure. As opposed to the similar operation of marginalization that comes from probability theory and is well-known in causal abstraction, the consolidation operation is capable of representing interventions on those variables that were abstracted since the consolidated variable is of a special kind. It also accommodates compressions since other compatible functions can be assigned to the consolidated variable; this also opens the possibility of widening the domains of the original variables.

Strengths

The problem is relatively well-situated within the literature, I believe the results to be sound, and the examples are good.

Weaknesses

I am not convinced of the utility of the approach. What open problems does this contribute to solving?

Questions

- The title seems inappropriate (especially imperative mood). Are you suggesting that consolidation is always superior to marginalization? - Line 4: "Thus, ... analyze." is not a full sentence - Lines 35--36: "Given that ... outcome." is not a full sentence. Awkward transition follows in "That is, ..." - Line 48: "intervention preserving" -> "intervention-preserving" - Line 69: it is inappropriate to cite do-calculus specifically here, rather, this is talking about the foundations of the SCM framework as a whole. - Line 116: "Section 4 four" -> "Section 4" - Is it really necessary to include the causal graph in the right of Figure 4?

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

2 fair

Contribution

2 fair

Limitations

Yes, they have.

Reviewer ZzBw2023-08-14

Thank you for answering my questions. I think I can understand most of these now. I wanted to make the following points: * Regarding the set of possible interventions: I think I see. If I do then there can be 0 or 1 unique intervention values v for each variable $X_i$ in some intervention set $I$, but the set of possible interventions $\curly{I}$ includes all possible values for each variable. Is that right? The set notation you use elsewhere is maybe more compact/readable than the equation in your response. * Regarding the "explicit computation" in Fig.2: I think I understand the explanation. Perhaps you could name the variable differently, e.g. $X_2'$ as it isn't the "same" variable, in the sense that it doesn't have the same properties as $X_2$. For instance, it may be the case that there is some original intervention $do(X_2=v)$, but this may not necessarily change the value of $X_2'$. Or am I misunderstanding? After reading the other reviews, the rebuttal and your responses to each of the reviewers I am raising my assessment to Accept.

Authorsrebuttal2023-08-15

Thank you for further engaging in the discussion, here are our comments to these two points: 1. Correct. The intervention set is a subset of the set of all possible interventions, where the set of all possible interventions is defined as the union over all intervenable variables and a union over each combination of values (from the respective domains of each of the variables) that these variables can take. We have switched to the following set notation: $\mathcal{I} \subseteq \\{\\{I_{i,v\_i}\\}\_{i \subseteq \\{1\dots N\\}}\\}\_{\mathbf{v} \in {\pmb{\mathcal{X}}}}$ where $v_i$ is the i-th element of $\mathbf{v}$ and such that $\mathbf{J} \subset \mathbf{I} \in \mathcal{I} \rightarrow \mathbf{J} \in \mathcal{I}$. (Reminder: $\pmb{\mathcal{X}}$ is defined as the cartesian product covering all variable domains $\mathcal{X}_i$.) Despite the rather minor technical drawback of generating some $\mathbf{I}$ 'multiple times' we appreciate the suggestion of a more compact formalization with reduced constrains. 2. We are sorry about the confusion regarding our answers. The consolidated $X_2$ will behave exactly the same (with and without interventions). This becomes more clear when looking only at the structural equations: $X_2$ is computed via $\rho_{\mathbf{E}'}(\mathbf{U}, \mathbf{I})$, which is required to be consistent with the original $\mathcal{M}$ (- compare to Def. 2). However, when considering Figure 2 without any additional precautions, a potential reader may get the impression that there are now two possible places to intervene: (1) via parameter $\mathbf{I}$, e.g. $\mathbf{I} = \\{ do(X_2 := v)\\}$, of $\rho_{\mathbf{E}'}(\mathbf{U}, \mathbf{I})$ as it computes the value of $X_2$ and (2) via a 'classical' intervention on the graph, $do(X_2 := v')$, cutting the edge between $X_2$ and $\rho_{\mathbf{E}'}$. This can not be, as two, possibly different, intervention values for $X_2$ would lead to an inconsistent behaviour. (Note that in the second case $\rho$ would not only compute an incorrect value for $X_2$, but also incorrectly compute values for the dependent $X_4$ and $X_5$). For this reason - and while observing the value of $X_2$ - we are not allowed to manipulate the connection between $\rho_{\mathbf{E}'}$ and $X_2$ (= we do not allow intervention of case (2)). To indicate this restriction in the graph, we decided for the altered visual representation. Thanks again for asking this. We will make sure to add the main points of this discussion in the figure caption. Thank you for helping in improving our paper by providing all of this great feedback! Also thank you for raising your score even further post-rebuttal to a clear accept. Very much appreciated and kind regards, your authors

Reviewer CKKq2023-08-14

I have read the response and will maintain my current rating.

Authorsrebuttal2023-08-16

Looking Forward to Feedback on Our Response

Dear Reviewer, We appreciate the time and effort that you have taken to provide us with the review. We would like to ask if the reviewer has any further concerns or is satisfied by our responses to the original review. We are looking forward to any further discussion with the reviewer and would like to thank the reviewer again for helping make our paper better. Regards, The Authors

Reviewer tWPL2023-08-21

Thanks very much for the clarification. I acknowledge that I have read the rebuttal. I will keep my score.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC