acceptodds
Under review as a conference paper at ICLR 2027

Causal Structure-Guided Optimal Transport for Treatment Effect Estimation

Abstract

Treatment effect estimation seeks to evaluate the effect of a treatment based on non-randomized trial data. To reduce the confounding bias between the control and treated groups, existing methods usually align the distributions of the groups by learning balanced representations from covariates. Despite their success, these methods simply model the covariates as a flat vector while ignoring the intrinsic causal structure among variables. To address this limitation, we exploit the causal structure among covariates to learn balanced representations for causal effect estimation. Specifically, we incorporate the causal structure into the framework of optimal transport, which has demonstrated a powerful ability for distribution alignment. We first discover the causal structure and establish the structural equations of covariates, and introduce the residuals into optimal transport, where the residuals capture individual information for heterogeneous treatment effect estimation. We further leverage the causal structure to refine the cost function in optimal transport, deriving a causal structure-aware transport plan. Theoretically, we reveal that the estimation error of the effect can be bounded by the proposed causal structure-guided optimal transport cost. Empirically, we conduct experiments on synthetic and real-world datasets to demonstrate the effectiveness of our proposed method.

Primary area:
causal reasoning
Keywords:
Causal Effect Estimation, Optimal Transport, Confounding Bias

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_7ouuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies CATE estimation from observational data, where two problems appear together: confounders make the treated and control groups differ, and the two groups' covariate distributions are shifted. The proposed method, CSOT, first runs causal discovery to recover a DAG and the structural equations, takes the residuals of these equations as structure-conditioned individual deviations, concatenates them with the covariates, and encodes [x,e][x,e] into a representation zz. It then builds the transport cost from two parts: the representation distance, and a residual distance where each covariate is weighted by a multi-hop structural influence sks_k computed from ∣A^∣+∣A^∣2|\hat A|+|\hat A|^2 and normalized to unit average. Entropic OT between the treated and control empirical distributions gives the balancing loss, and the total objective is just the factual loss plus the OT loss. The paper proves a balancing error bound, a PEHE bound in terms of the structure-guided Wasserstein discrepancy, and a finite-sample version. Experiments are on IHDP, ACIC 2016 and a self-constructed synthetic benchmark, and CSOT reports the best within-sample εPEHE\sqrt{\varepsilon_{\text{PEHE}}} on IHDP (0.552) and ACIC (1.277).

Critical strengths
  • Combining causal structure with OT is a reasonable idea. OT-based CATE methods do generally ignore the dependence structure among covariates, and structure-based methods generally ignore the distribution shift, so using multi-hop influence weights to shape the transport cost is a natural direction.
  • The three theorems form a complete chain: a balancing error bound, a PEHE bound, and a finite-sample bound. The mathematics is correct under the stated assumptions and is consistent with the classical CFR-Wass framework.
Critical weaknesses
  • The theory is agnostic about the causal structure. The balancing bound holds for any cost function, so no theorem says that the structure-guided cost is better than the Euclidean one; what actually supports the method is the experiments, and the experimental controls are not clean enough for that. In particular, there is no random-graph or constant-weight control, which would show whether the gain comes from the structure itself or just from having some weighted cost, and no ablation that swaps the structure-guided cost for CFR-Pro's proximity cost while holding the backbone fixed.
  • Under a linear SEM the residual is a linear transformation of the covariates. The structural equations are fitted by linear ridge regression, so e=(I−W^)xe=(I-\widehat W)x and [x,e][x,e] contains exactly the same information as xx; the encoder could learn this transformation on its own. The claim that the residuals carry individual information that is otherwise unavailable therefore does not hold at the level of the implementation, and it would be worth either using a nonlinear fjf_j or arguing explicitly that the benefit is an inductive bias rather than new information.
Questions for authors
  • What is the name "SROT-CFR" in Appendix C.1 means.
  • How does CSOT behave when the true structure is nonlinear? The synthetic DGP already contains nonlinear/interaction terms, so please report a recovered-graph-quality diagnostic (e.g. SHD to the true DAG) and a DAG-misspecification sensitivity study, to show the linear discovery is good enough.
2 comments

Weakness 1. Causal structure in the theory and random-graph, constant-weight, and fixed-backbone cost controls.

Answer:

1. The causal structure enters Theorem 1 through the cost

Cij2=∥z1,i−z0,j∥22+λe∑k=1dsk(e1,ik−e0,jk)2.C_{ij}^{2} = \lVert z_{1,i}-z_{0,j}\rVert_{2}^{2} + \lambda_e\sum_{k=1}^{d}s_k(e_{1,ik}-e_{0,jk})^{2}.

The estimated graph determines both the structural residuals and their two-hop weights sks_k, influencing cross-group alignment. When the squared prediction losses satisfy the condition of Theorem 1, W(C,P1,P0)W(C,P_1,P_0) bounds their cross-group differences, connecting this structure-guided alignment to treatment effect estimation error.

2. On synthetic data (α=γ=1\alpha=\gamma=1, ten data seeds), CSOT uses DirectLiNGAM. For the random-graph control, we randomly order the covariates and sample 32 forward edges, matching the true DAG's edge count and ensuring acyclicity. We refit the residuals and recompute sks_k using each random graph.

SettingϵPEHE↓\sqrt{\epsilon_{\mathrm{PEHE}}}\downarrow
CSOT0.1240±0.01830.1240\pm0.0183
sk=1s_k=10.1308±0.01420.1308\pm0.0142
Random graph0.1234±0.0195\boldsymbol{0.1234\pm0.0195}

Graph-based weights reduce mean error by 5.2% relative to constant weights. CSOT also maintains comparable prediction accuracy under the random-graph control, suggesting robustness to graph misspecification in this setting.

3. We also replace the structure-guided cost with our implementation of CFR-Pro's proximity cost, keeping the backbone fixed, on ACIC experiment 1.

CostϵPEHE↓\sqrt{\epsilon_{\mathrm{PEHE}}}\downarrow
CSOT1.0989\boldsymbol{1.0989}
CFR-Pro1.14801.1480

With the backbone fixed, the structure-guided cost achieves 4.3% lower test error than CFR-Pro's proximity cost. This controlled comparison supports the contribution of our cost design to improved treatment effect estimation in this experiment.

Weakness 2. Structural residuals as an inductive bias rather than additional information.

Answer:

We thank the reviewer for this observation. We agree that, once the linear structural equations are fitted, ee is determined by xx. As described in Section 4.1, the individual information captured by a structural residual refers to the variation in a covariate that is unexplained by its estimated parents. Explicitly incorporating these residuals into the encoder input and transport cost emphasizes such individual variations, helping prevent them from being overlooked during representation learning and cross-group alignment. This introduces a structure-aware inductive bias.

Q1: What is the name "SROT-CFR" in Appendix C.1 means.

Answer:

We apologize for this oversight. The name “SROT-CFR” in Appendix C.1 is incorrect and should read “CSOT”. We will correct it and check the manuscript for consistent method naming throughout.

Question 2. Nonlinear structures, graph-recovery diagnostics, and DAG-misspecification sensitivity.

Answer:

We thank the reviewer for this question.

1. Nonlinearity and graph quality.

We evaluate CSOT with DirectLiNGAM on synthetic data with α=1\alpha=1 under different nonlinearity strengths γ\gamma. As shown in Table 9, CSOT achieves mean test errors of 0.12400.1240 and 0.12300.1230 at γ=1\gamma=1 and 22, respectively, compared with 0.12470.1247 under linear covariate mechanisms (γ=0\gamma=0). The table also reports the SHD to the true DAG.

γ\gammaSHD ↓\downarrowTest ϵPEHE\sqrt{\epsilon_{\mathrm{PEHE}}} ↓\downarrow
0021.5±2.17\boldsymbol{21.5\pm2.17}0.1247±0.01470.1247\pm0.0147
0.50.522.6±3.8622.6\pm3.860.1354±0.02240.1354\pm0.0224
1122.3±3.2022.3\pm3.200.1240±0.01830.1240\pm0.0183
2224.3±2.4124.3\pm2.410.1230±0.0134\boldsymbol{0.1230\pm0.0134}

Although SHD increases from 21.521.5 at γ=0\gamma=0 to 24.324.3 at γ=2\gamma=2, CSOT maintains a comparable mean prediction error (0.12470.1247 versus 0.12300.1230). These results support the effectiveness of CSOT with linear discovery under the nonlinear causal structures tested here.

2. Graph misspecification.

We compare the true graph, the estimated graph, and perturbed graphs on synthetic data; the results are shown in the table below.

GraphSHD ↓\downarrowTest ϵPEHE\sqrt{\epsilon_{\mathrm{PEHE}}} ↓\downarrow
True graph000.1174±0.0142\boldsymbol{0.1174\pm0.0142}
Estimated graph (DirectLiNGAM)22.322.30.1240±0.01830.1240\pm0.0183
Perturbed graph (10%)660.1208±0.01830.1208\pm0.0183
Perturbed graph (20%)12120.1245±0.02190.1245\pm0.0219
Perturbed graph (40%)26260.1180±0.01030.1180\pm0.0103

The true graph achieves the lowest mean error, while the estimated and perturbed graphs retain comparable prediction performance. Even at 40% perturbation, the mean error remains close to the true-graph result (0.11800.1180 versus 0.11740.1174), supporting the robustness of CSOT to the graph misspecifications tested here.

Dear Reviewer 7ouu,

Thank you for your comments. We would welcome your feedback on the added experimental controls and clarification of structural residuals. If our responses have addressed your concerns, we would greatly appreciate your support for our paper during the reviewer–AC discussion.

Sincerely,

The authors

Reviewer_DFMeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Causal Structure-Guided Optimal Transport (CSOT) for treatment effect estimation. It combines causal discovery, structural residuals, and graph-based weighting to improve representation balancing. Experiments on IHDP, ACIC 2016, and synthetic datasets demonstrate improved estimation accuracy over existing methods.

Critical strengths
  1. The paper explores an interesting approach to integrating causal structure into optimal transport.
  2. The proposed framework is clearly presented and technically well organized.
  3. The experiments show promising results across multiple datasets, supported by ablation and sensitivity analyses.
Critical weaknesses

Weaknesses

  1. The accuracy of the discovered causal structures is not sufficiently validated (Section 4.1). Comparisons with ground-truth and perturbed graphs are needed.

  2. The contribution of structural residuals is unclear (Section 4.1). Comparisons with standard regression residuals and feature augmentation would help establish their effectiveness.

  3. The graph-based weighting strategy lacks sufficient justification (Section 4.2). Comparisons with alternative weighting methods are needed.

  4. The theoretical analysis contains questionable derivations (Section 4.4 and Appendix B), including an invalid inequality in Equation (41). The proofs and assumptions require revision.

  5. The experimental evaluation remains limited (Section 5). Additional tests on nonlinear causal structures, computational efficiency, and reproducibility are needed.

Questions for authors

see weaknesses

3 comments

Weakness 1. Comparisons with ground-truth and perturbed graphs.

Answer:

We thank the reviewer for this suggestion. We evaluate the graph discovered by DirectLiNGAM against the true DAG and compare treatment effect estimation using the true, estimated, and perturbed graphs on synthetic data, with results shown in the tables below.

GraphTest ϵPEHE\sqrt{\epsilon_{\mathrm{PEHE}}} ↓\downarrow
True graph0.1174±0.0142\boldsymbol{0.1174\pm0.0142}
Estimated graph (DirectLiNGAM)0.1240±0.01830.1240\pm0.0183
Perturbed graph (10%)0.1208±0.01830.1208\pm0.0183
Perturbed graph (20%)0.1245±0.02190.1245\pm0.0219
Perturbed graph (40%)0.1180±0.01030.1180\pm0.0103

The true graph achieves the lowest mean error (0.11740.1174), while the estimated and perturbed graphs yield errors between 0.11800.1180 and 0.12450.1245. These results suggest that CSOT benefits from the true graph and maintains stable treatment effect estimation under the graph misspecifications tested here.

Weakness 2. Structural residuals versus standard regression residuals and feature augmentation.

Answer:

We compare structural residuals with standard regression residuals and feature augmentation on synthetic data with α=1\alpha=1, using DirectLiNGAM. Here, eirege_i^{\mathrm{reg}} denotes the residuals from regressing each covariate on all other covariates, x~i\widetilde{x}_i denotes standardized covariates, and RR is a fixed Gaussian random-projection matrix; the results are shown in the table below.

Input uiu_iTest ϵPEHE\sqrt{\epsilon_{\mathrm{PEHE}}} ↓\downarrow
[xi,ei][x_i,e_i]0.1240±0.0183\boldsymbol{0.1240\pm0.0183}
[xi,eireg][x_i,e_i^{\mathrm{reg}}]0.1381±0.02800.1381\pm0.0280
[xi,x~iR][x_i,\widetilde{x}_iR]0.1257±0.01590.1257\pm0.0159
[xi,x~i][x_i,\widetilde{x}_i]0.1254±0.02230.1254\pm0.0223
[xi,0][x_i,\mathbf{0}]0.1247±0.01290.1247\pm0.0129

Structural residuals achieve the lowest mean error, with a reduction of approximately 10.2%10.2\% compared with standard regression residuals. These results support the effectiveness of structural residuals for treatment effect estimation in this setting.

Weakness 3. Alternative weighting methods and justification of graph-based weights.

Answer:

We thank the reviewer for this suggestion. Our weighting strategy uses the one-hop and two-hop influence of each covariate in the estimated graph, where B=∣A^X∣B=|\widehat A_X|. We compare it with uniform, one-hop, and incoming two-hop weights while keeping the other model components and training settings fixed. The scores are normalized to have approximately unit mean. Table 4 reports results on synthetic data with α=1\alpha=1, using DirectLiNGAM. The results are mean ±\pm standard deviation over ten data seeds.

s~k\widetilde{s}_kTest ϵPEHE\sqrt{\epsilon_{\mathrm{PEHE}}} ↓\downarrow
∑j=1d[Bkj+(B2)kj]\sum_{j=1}^{d}\left[B_{kj}+(B^2)_{kj}\right]0.1240±0.0183\boldsymbol{0.1240\pm0.0183}
110.1308±0.01420.1308\pm0.0142
∑j=1dBkj\sum_{j=1}^{d}B_{kj}0.1255±0.02670.1255\pm0.0267
∑j=1d[Bjk+(B2)jk]\sum_{j=1}^{d}\left[B_{jk}+(B^2)_{jk}\right]0.1268±0.01240.1268\pm0.0124

The proposed two-hop weights achieve the lowest mean error among these alternatives, reducing it by approximately 5.2%5.2\% relative to uniform weighting. These results support the effectiveness of the proposed graph-based weighting strategy in this setting.

Weakness 4. Correcting Equation (41), the proofs, and the assumptions.

Answer:

We thank the reviewer for pointing out these issues.

The original proof omitted the factor of 22 in the squared inequality and the group probabilities when expanding the population distribution. We will correct the proof. When the two squared prediction losses satisfy the conditions of Theorem 1, the bound becomes

ϵPEHE≤2E1(1)+2E0(0)+O(W(C,P1,P0)).\epsilon_{\mathrm{PEHE}} \leq 2\mathcal E_1^{(1)} + 2\mathcal E_0^{(0)} + \mathcal O\left(\mathcal W(C,P_1,P_0)\right).

Since Theorem 3 relies on Theorem 2, we will rederive its proof using the corrected result and verify the conditions of the finite-sample transport bound and the sample sizes of both groups.

Weakness 5. Nonlinear causal structures, computational efficiency, and reproducibility.

Answer:

1. Nonlinear causal structures.

We thank the reviewer for these suggestions. We vary the nonlinearity strength γ\gamma in the synthetic covariate mechanisms. For example,

X10=0.6[X9+γ(tanh⁡(X9)−X9)]+ε10.X_{10}=0.6\big[X_9+\gamma(\tanh(X_9)-X_9)\big]+\varepsilon_{10}.

Here, γ=0\gamma=0 gives linear covariate mechanisms, γ=1\gamma=1 recovers the original mechanisms, and γ=2\gamma=2 amplifies their departure from the linear mechanisms. We will include the complete data-generation mechanisms in the appendix of the revised paper. Results for CSOT with DirectLiNGAM and NOTEARS at α=1\alpha=1 are shown below.

γ\gammaCSOT-DirectCSOT-NOTEARS
000.1247±0.0147\boldsymbol{0.1247\pm0.0147}0.1314±0.01280.1314\pm0.0128
0.50.50.1354±0.02240.1354\pm0.02240.1280±0.0194\boldsymbol{0.1280\pm0.0194}
110.1240±0.0183\boldsymbol{0.1240\pm0.0183}0.1329±0.02300.1329\pm0.0230
220.1230±0.0134\boldsymbol{0.1230\pm0.0134}0.1252±0.01340.1252\pm0.0134

CSOT maintains treatment effect estimation accuracy under the tested nonlinear causal structures.

2. Computational efficiency.

The runtime results on an NVIDIA RTX 4090 are shown in the table below.

MethodRuntime (s)
CSOT-Direct217.93217.93
CSOT-ICA134.70134.70
CSOT-NOTEARS132.14132.14
CSOT-PC190.34190.34
CSOT-GOLEM133.64133.64

With 68.01%68.01\% of samples having true propensity scores below 0.10.1 or above 0.90.9, CSOT achieves mean test errors of 0.15400.1540--0.16930.1693 across five causal discovery methods. These results support the applicability of CSOT under the weak-overlap setting tested here.

3. Reproducibility.

We train CSOT using Adam with a learning rate of 10−310^{-3}, batch size 128128, and 200200 epochs. The hidden and representation dimensions are 200200 and 6464. Detailed implementation configurations will be provided in the appendix, and the code and experiment configurations will be released upon acceptance.

Dear Reviewer DFMe,

Thank you for your suggestions. We would welcome your feedback on the added graph, residual, and weight comparisons. If our responses have addressed your concerns, we would greatly appreciate your support for our paper during the reviewer–AC discussion.

Sincerely,

The authors

Reviewer_j22UOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper considers the estimation of the conditional average treatment effect (CATE). It proposes to residualize the covariates according to an estimated DAG and use the residuals to balance between the control and treatment arms when estimating CATE. Theoretical results and numerical studies are provided.

Critical strengths

NA

Critical weaknesses
  1. Theoretical results. Some of the key theoretical results are incorrect. For example, Theorem 2: with identical treated and control distributions, true outcomes equal to zero, and predictions equal to 1 and -1, the squared treatment effect error is 4, while the claimed upper bound is 2. Theorem 3 has similar problems. Furthermore, the theorems do not sufficiently support the method. For example, Theorem 1 is a generic transport inequality that does not depend on causal structure. The analysis does not establish why the proposed residual distances and graph based weights improve estimation. In addition, most theorems appear to provide decompositions of error sources, without much deeper insight.

  2. Method. First, the causal DAG among the predictors may not be known, and learning it from the data is challenging. In addition, the method encourages alignment between the control and treatment groups using the residuals. This may not always be feasible or appropriate, especially under weak/no overlap. The weighting scheme is also somewhat counterintuitive. For example, it assigns zero residual weight to a covariate with no outgoing edges, even if that covariate strongly predicts the outcome.

  3. Presentation. The presentation primarily describes the concrete steps but provides little intuition behind them. For example, the paper does not sufficiently explain why similarity in structural residuals implies useful counterfactual comparability.

Questions for authors

My questions are related to the weaknesses.

  1. Could the authors correct the proofs and align the theoretical results more closely with the proposed method?

  2. Could the authors provide more intuition behind the method?

4 comments

Weakness 1 & Question 1. Theoretical correctness and the connection to causal structure.

Answer:

1. We thank the reviewer for pointing out the errors in Theorems 2 and 3. The original proof omitted the factor of 22 in the squared inequality and the group probabilities when expanding the population distribution. We will correct the proof. When the two squared prediction losses satisfy the conditions of Theorem 1, the bound becomes

ϵPEHE≤2E1(1)+2E0(0)+O(W(C,P1,P0)).\epsilon_{\mathrm{PEHE}} \leq 2\mathcal E_1^{(1)} + 2\mathcal E_0^{(0)} + \mathcal O\left(\mathcal W(C,P_1,P_0)\right).

Combining the corrected Theorem 2 with the finite-sample transport bound in Lemma 4, under its stated conditions, gives the following result with probability at least 1−2δ1-2\delta:

ϵPEHE≤2E1(1)+2E0(0)+O(W(C,P^1,P^0))+O(1δn).\epsilon_{\mathrm{PEHE}} \leq 2\mathcal E_1^{(1)} + 2\mathcal E_0^{(0)} + \mathcal O\left(W(C,\widehat P_1,\widehat P_0)\right) + \mathcal O\left(\frac{1}{\sqrt{\delta n}}\right).

2. We thank the reviewer for this question in the theoretical results relate to the causal structure. The estimated graph determines both the structural residuals and their weights sks_k in the transport cost CC,

Cij2=∥z1,i−z0,j∥22+λe∑k=1dsk(e1,ik−e0,jk)2.C_{ij}^{2} = \lVert z_{1,i}-z_{0,j}\rVert_{2}^{2} + \lambda_e\sum_{k=1}^{d}s_k(e_{1,ik}-e_{0,jk})^{2}.

Thus, even when two pairs of units have the same representation distance, their transport costs can differ because of their structural residual differences. The resulting transport plan penalizes alignment across large structural deviations, particularly on covariates with greater two-hop influence. When the squared prediction losses satisfy the condition of Theorem 1, W(C,P1,P0)W(C,P_1,P_0) bounds their cross-group differences, connecting the factual prediction errors to the treatment effect estimation error.

Weakness 2. Unknown DAGs, weak/no overlap, and zero residual weights.

Answer:

1. Causal graph discovery. We use existing causal discovery methods to estimate the covariate graph from data, so the true DAG does not need to be known in advance.

2. We evaluate CSOT with DirectLiNGAM under weak overlap on synthetic data and report the results

Methodα\alphaExtreme propensityOut ϵPEHE\sqrt{\epsilon_{\mathrm{PEHE}}} ↓\downarrow
CSOT-Direct101068.01%68.01\%0.1693±0.03580.1693\pm0.0358
CSOT-ICA101068.01%68.01\%0.1659±0.03220.1659\pm0.0322
CSOT-NOTEARS101068.01%68.01\%0.1540±0.0257\boldsymbol{0.1540\pm0.0257}
CSOT-PC101068.01%68.01\%0.1613±0.02310.1613\pm0.0231
CSOT-GOLEM101068.01%68.01\%0.1591±0.02760.1591\pm0.0276

With 68.01%68.01\% of samples having true propensity scores below 0.10.1 or above 0.90.9, CSOT achieves mean test errors of 0.15400.1540--0.16930.1693 across five causal discovery methods. These results support the applicability of CSOT under the weak-overlap setting tested here.

3. The weight sks_k is constructed from the one-hop and two-hop influence of XkX_k in the covariate graph. A variable without outgoing edges has sk=0s_k=0, although it may still predict the outcome. Its covariate value and residual remain in the encoder input and can still influence representation learning and outcome prediction.

Weakness 3 & Question 2. Intuition behind structural residuals and counterfactual comparability.

Answer:

We thank the reviewer for the suggestion. The intuition behind CSOT is to align treated and control units using both their representations and their deviations from the estimated structural relationships. Specifically, structural residuals measure how much each covariate deviates from the value predicted by its estimated parent variables. By incorporating these residuals into representation learning and the transport cost, CSOT encourages cross-group alignment that also takes these deviations into account. The graph-based weights further reflect the different structural roles of the covariates.

I thank the authors for the reply and correcting the proofs.

Overall, the procedure requires the knowledge of the graph, as well as learning the joint distribution of the residuals for matching, which could be computationally challenging (especially when the data is of medium size). As a result, I will maintain my score.

We thank the reviewer for acknowledging the corrected proofs. We would like to clarify the computational requirements.

1. Knowledge of the graph. CSOT estimates the covariate graph from the training data using existing causal discovery methods. Graph discovery, structural regression, and the computation of residuals and graph-based weights are performed once before representation learning.

2. Joint distribution of the residuals. CSOT uses the empirical distribution of residual vectors and computes their pairwise weighted distances directly. It does not fit a separate joint density model.

3. Computational cost. We measure runtime for CSOT with five causal discovery methods on ACIC 2016 (4,802 samples, experiment 1), using 200 epochs, batch size 128, and an NVIDIA RTX 4090. The results are shown below.

MethodRuntime (s)
CSOT-Direct217.93
CSOT-ICA134.70
CSOT-NOTEARS132.14
CSOT-PC190.34
CSOT-GOLEM133.64

These results show that CSOT runs efficiently on ACIC 2016.

Dear Reviewer j22U,

Thank you for your follow-up comments. We would welcome your feedback on our additional explanations and runtime results. If our responses have addressed your concerns, we would greatly appreciate your support for our paper during the reviewer–AC discussion.

Sincerely,

The authors

Comments on the paper

1 comment

Dear AC,

We appreciate the reviewers' comments, which help us improve the manuscript. We summarize the reviews and responses as follows.

1. We sincerely thank the reviewers for their positive evaluations of our submission.

Reviewer DFMe:

"The paper explores an interesting approach to integrating causal structure into optimal transport."

"The proposed framework is clearly presented and technically well organized."

"The experiments show promising results across multiple datasets, supported by ablation and sensitivity analyses."

Reviewer 7ouu:

"Combining causal structure with OT is a reasonable idea. OT-based CATE methods do generally ignore the dependence structure among covariates, and structure-based methods do generally ignore the distribution shift, so using multi-hop influence weights to shape the transport cost is a natural direction."

2. We provide detailed responses in the rebuttal and summarize the corresponding manuscript revisions below.

Reviewer j22U:

(1) For theoretical correctness and the connection to causal structure (Weakness 1 and Question 1), we revise the bounds in Theorems 2 and 3 and explain how the estimated graph determines the structural residuals and weights in the transport cost, shaping the transport plan and cross-group alignment. The corresponding revisions are in Section 4.4 and Appendix B.

(2) For unknown DAGs, weak/no overlap, and zero residual weights (Weakness 2), we describe graph estimation in Section 4.1 and explain in Section 4.2 how zero-weight covariates can still influence representation learning and outcome prediction. Weak-overlap experiments are reported in Appendix C.5.

(3) For the intuition behind the method (Weakness 3 and Question 2), we explain how structural residuals emphasize individual variations and guide cross-group alignment in Sections 4.1 and 4.2.

Reviewer DFMe:

(1) For graph-recovery accuracy (Weakness 1), we report SHD to the true DAG and compare CSOT using true, estimated, and perturbed graphs in Appendix C.6, Graph Perturbation.

(2) For the contribution of structural residuals (Weakness 2), we compare them with standard regression residuals and feature augmentation in the Input Ablation part of Appendix C.3.

(3) For the graph-based weighting strategy (Weakness 3), we compare the proposed two-hop weights with uniform, one-hop, and incoming two-hop weights in the Weight Ablation part of Appendix C.3.

(4) For the theoretical derivations (Weakness 4), we correct the missing factor of 22 in the squared inequality and update the corresponding bounds in Section 4.4 and Appendix B.2–B.3.

(5) For nonlinear causal structures, computational efficiency, and reproducibility (Weakness 5), we report nonlinear results in Section 5.1 and discuss them in Section 5.2, with the nonlinear mechanisms described in Appendix C.2. Runtime measurements and experimental configurations are provided in Appendix C.7. The source code, complete experiment configurations, and software environment requirements will be released upon acceptance.

Reviewer 7ouu:

(1) For causal structure in the theory and experimental controls (Weakness 1), we explain the role of the structure-guided cost in Section 4.4 and provide random-graph, constant-weight, and fixed-backbone CFR-Pro cost comparisons in the Graph and Weight Controls and Cost Ablation parts of Appendix C.3.

(2) For the information carried by structural residuals (Weakness 2), we clarify in Section 4.1 that fitted residuals are determined by the covariates and explain how explicitly incorporating them emphasizes individual variations and introduces a structure-aware inductive bias.

(3) For method naming (Question 1), we correct "SROT-CFR" to "CSOT" in Appendix C.

(4) For nonlinear structures and DAG misspecification (Question 2), we report nonlinear prediction results in Sections 5.1–5.2 and describe the nonlinear mechanisms in Appendix C.2. SHD to the true DAG and graph-misspecification sensitivity results are reported in Appendix C.6.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.