__ Summary and Contributions__: The paper transforms the identifiable causal inference problem as a weighted empirical risk minimization problem. It proposes a new objective for WERM and a learning algorithm based on it. The simulation shows WERM-ID achieves low absolute error and computational time.

__ Strengths__: The paper provides a general framework to cast the causal estimand to a form of weighted ERM. The method is widely applicable to a variety of causal graphs. This paper contains rich materials. Viewing causal inference from an optimization perspective is relevant to the NeurIPS community.

__ Weaknesses__: 1. Under the identification assumption, if other causal functionals exist, such as regression adjustment and IPTW, what is the advantage of the WERM-ID?
2. In the empirical simulation, there proposed method is only compared with plug-in estimands. What about other weighting methods?
3. The paper contains many technical details but lacks a detailed explanation. If the page limit is a major concern, the author may need to decide what are the essential messages to express and explain them very clearly, instead of thrusting and skipping over many materials. See the Clarity section for more details.
######
post rebuttal comment:
The author's feedback addresses the major concerns of the reviewer. I modified the evaluation accordingly to marginal above acceptance.

__ Correctness__: The paper claims the algorithm can “generate ANY identiﬁable causal functionals as weighted distributions”. This claim excludes any counterexample; is it too strong? Is the identiﬁcation as the only assumption that is needed for WERM?

__ Clarity__: The explaining example is not very clear. Some formulations need further explanation:
i) Why is Eq. (1) true?
ii) What does the dash curve with double arrow in Figure 1 mean? It is not a standard notation in DAG. Is there a cycle X -> W ->R ->X?
iii) Why p(y | do(x)) = p(y | do(r), x)?
iv) There are also many skipped steps in the derivation above Lemma 1 for P(y|do(x)) = P (x, y|do(r)) /P (x|do(r)) = P(y|do(r), x) = P^W(y|x, r)

__ Relation to Prior Work__: The connection with previous weighting methods is not well explained.

__ Reproducibility__: Yes

__ Additional Feedback__:

__ Summary and Contributions__: The paper tackles the problem of assessing the strength of the causal effect in the presence of unobserved common causes from observational data when the assumed causal relationships among the observed variables are given in the form of a "causal" graph. In this, the paper extends the work of "On the identification of causal effects" [48] by formulating the problem of learning functionals as a weighted ERM problem.

__ Strengths__: A polynomial-time algorithm that is able to learn from finite samples that improve the performance over the proposed baseline. Elaborate derivations and presumably correct proofs (I did not check the proofs).

__ Weaknesses__: The paper is, somewhat surprisingly, is extremely difficult to follow. It is written in some form of nonlinear fashion with thoughts and references jumping, at times, back across a few pages. Although based on [48], this paper is much harder to follow unlike it, and is lacking the context of what and why is being done. In this sense, it is hardly self-contained. Potentially, the jargon and context may be clear to narrow specialists working precisely in this topic, but I doubt a wider machine learning readership, even well versed in causal learning, will find the paper easy to follow. I did invest substantial time and only after going through [48] some things became clearer, but not all.
The lack of clarity could have been attributed to the page limit and the needed to be dense. However, the introduction practically occupies 3 pages out of 8 and doesn't explain that the causal graph needs to be provided beforehand and we only need this new method because of unobserved common causes, since otherwise the causal graph and P(V) completely define everything.
In general, the paper feels like a good fit for a journal, where it can be safely combined with the supplemented proofs and additional experiments on more realistic graphs larger than the 3-4 variable examples could be given. The intended readership could be statisticians and practical relevance to any graphs of interesting scale is irrelevant, I leave that as a possibility.
The example graphs are unrealistic for interesting practical applications in ML. The polynomial-time algorithm runs in close to a few seconds if not faster, and yet no empirical evidence of performance is given on large graphs. Potentially, sample complexity may be exponential in the node number and the algorithm can only handle 3-4 nodes. This is not clear from the paper.

__ Correctness__: I did not check the proofs in the supplement. Empirical validation is extremely limited as explained above.

__ Clarity__: Either written for a very narrow audience or is just poorly written, overwhelming rather than explaining.

__ Relation to Prior Work__: Relevant literature is discussed

__ Reproducibility__: No

__ Additional Feedback__:

__ Summary and Contributions__: In this paper, the authors provide an approach for estimating causal effects by weighted predictors. It is a good try to causal estimation. Surprisingly, it is revealed that causal effects representation can be converted to weighted distributions.

__ Strengths__: The paper provides solid theory about converting the representation of identified causal effects to the format of weighted distribution. Moreover, under the assumption that W*\in \mathcal{H}_W, the authors prove the consistency and converge rate of the estimated causal effect.

__ Weaknesses__: I guess it is a bit not friendly to readers who are not quite clear about this region.

__ Correctness__: They seem right. But I did not check them throughly.

__ Clarity__: Yes. But it is a bit difficult to read.

__ Relation to Prior Work__: Yes.

__ Reproducibility__: Yes

__ Additional Feedback__: I am a bit curious about the comparison results with some recent causal inference methods under PO framework if simply seeing the whole other variables (V\{X,Y}) as observed confounders. I know that PO framework does not take the causal graph into account and it is not reasonable from the viewpoint of identifiability in this setting. I am just curious about the performance gap between the proposed method and the SOTA in related regions. Because although the method in this paper is theoretically reasonable, more errors may be introduced due to the complicated computation (such as estimating many conditional possibilities). So I have this question. By the comparison, at least it can tell which present method for estimating causal effect is better for a user without much related knowledge but just wants to estimate the causal effect. I totally believe that the framework in this paper has unlimited development possibilities and could have better and better performance in the long run. This paper is a good start. So the comparison results will not influence my judgment on this paper.
Supplementary Page 279: Should "Thm. 1" be "Prop. 1"?

__ Summary and Contributions__: This paper provides a learning framework for causal inference when the conditional ignorability assumption does not hold although the causal effect is identifiable. The framework provides an algorithm to first massage the desired estimand (using weighted distributions) to look like a back-door expression (on which we can apply the conventional Empirical Risk Minimization (ERM) technique) and then learn the required weights on the previous step to complete the training procedure. The algorithm works in an end-to-end fashion.

__ Strengths__: Parts of the claimed contributions of this work are significant and to the best of my knowledge, has not been explored before.
Causal inference is a relevant subject to the NeurIPS community.

__ Weaknesses__: I have many questions regarding the soundness of the claims; especially theoretical grounding. I have elaborated on them in the “correctness” section.

__ Correctness__: Eq. (1): What happened to the variable “r”?
Lines 89-90: Why does “P(x, y, w | do(r))” equal “1/P(r | w) * P(x, y, w, r)”?
Algorithm 1, line 8.5: What is “T”?
Learning low variance weights is not novel as (Swaminathan and Joachims, 2015) have already addressed it in their Counterfactual Risk Minimization (CRM) framework. Please comment on whether/how your work differs from theirs?
Swaminathan, A., & Joachims, T. (2015). Counterfactual risk minimization: Learning from logged bandit feedback. In International Conference on Machine Learning (ICML).

__ Clarity__: Use of too many inline mathematical statements (some even essential to understanding an idea) has made it difficult to keep the reading flow.

__ Relation to Prior Work__: There is a nice selection of literature reviewed in the paper; however, there is no explicit discussion on how this work differs from the prior work.

__ Reproducibility__: No

__ Additional Feedback__: Please address my comments in the rebuttal and I shall update my score accordingly.
===== post-rebuttal comments =====
I appreciate the authors' response to my questions. Although these cover my concerns, I think they should incorporate them into the final manuscript. I have updated my score to 6.