acceptodds
Under review as a conference paper at ICLR 2027

Geometric Mutual Information for Causal Effect Estimation under Network Interference

Abstract

Estimating causal effect on network data considers the situation with the presence of dependencies between individuals. To reduce complex confounding bias caused by network interference, existing methods usually adopt the same information aggregation function for covariates and treatments. However, these methods fail to exploit different geometric structures in the spaces of covariates and treatments, which are vital for information aggregation on network data. In this paper, starting from the causal effect estimation error, we propose a geometric mutual information model to learn balanced representations, so that different geometric structures involved in covariates and treatments are leveraged in representation learning and information aggregation. Specifically, we theoretically reveal that balanced representation learning on network data can be achieved by maximizing the dissimilarity between the geometric structures of covariates and treatments measured by the Gromov-Wasserstein discrepancy. To better characterize the geometry of the network data, we apply the Ricci flow to refine the graph structures for representation learning as well as information aggregation. We conduct experiments on benchmark datasets to demonstrate the effectiveness of our method.

Primary area:
causal reasoning
Keywords:
Causal Effect Estimation, Network Data, Balanced Representation Learning, Gromov-Wasserstein Discrepancy

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

The paper proposes GEOMINI for causal effect estimation under network interference. It constructs separate geometries for representations and treatments, uses Ricci flow to obtain neighborhood aggregation weights, and trains with factual prediction loss and a negative Gromov-Wasserstein discrepancy. Mutual-information bounds motivate the objective. Evaluation covers semi-synthetic BlogCatalog and Flickr data, with an additional power-plant application.

Critical strengths
  1. Using separate geometries for representations and treatments is a well-motivated design. It allows neighbor features and neighbor treatments to receive different aggregation weights, with an explicit account of how these weights enter outcome prediction.

  2. The experiments include several network causal estimators, with particularly competitive direct-effect results. The appendices also examine treatment flips, hyperparameter sensitivity, the omitted objective term, a real-data application, and runtime. These provide useful evidence about practical behavior.

Critical weaknesses
  1. There is a gap between the theory and representation learning. Theorem 4.2 assumes an invertible mapping, and Eq. (29) explicitly makes the true mutual information equal to that of the original covariates. It therefore cannot decrease through learning an invertible representation. The paper does not specify an invertibility guarantee for the implemented MLP. I could not see how this bound supports the proposed treatment-invariant representation learning. This concerns the link to training, rather than the derivation of Theorem 4.2 under its assumptions.

  2. The representation-dependent Frobenius term in Eq. (7) is omitted from training. Its 1/N² coefficient alone does not make it negligible, since its squared Frobenius norm sums N² squared pairwise discrepancies and the normalized quantity can remain of constant order. I appreciate the ablation in Appendix I, which provides practical evidence in one setting, but it does not establish that maximizing GW alone minimizes the stated upper bound.

  3. Some equations could be combined to make the paper easier to read. For example, Eqs. (12) and (13) could be written as one formula. Cutting this repetition would help readers follow the main argument.

Questions for authors
  1. Can the analysis cover the actual encoder and specify a condition that controls the omitted matrix term during optimization?

  2. Equation (38) in Appendix D.3 uses nonnegative conditional entropy, while the subsequent argument uses densities and Jacobians for continuous variables. What additional assumption justifies this step, or how should the lemma be amended?

  3. Equation (77) uses different exposure levels for the estimate and ground truth. Appendix J also calls f(1,z)−f(0,z) a total effect, although the earlier definitions make this a direct-effect contrast at fixed exposure. Are these notation errors, and which contrasts were used to produce the results?

4 comments

W1/Q2: The theory relies on invalid invertibility and nonnegativity assumptions.

A: We thank the reviewer for identifying this issue. We agree that the change-of-variables argument in Eqs. (29) and (42)–(44) unnecessarily relies on the invertibility and nonnegative of ϕ\phi and is not needed in our discrete setting.

(1) In Eq. (29), since (ϕ(X),ϕ(XN))(\phi(X),\phi(X_{\mathcal N})) is determined by (X,XN)(X,X_{\mathcal N}), we have \begin{align} (\phi(X),\phi(X_{\mathcal N})) \mathrel{\perp\perp} (T,Z)\mid (X,X_{\mathcal N}), \end{align} and hence \begin{align} I\left((\phi(X),\phi(X_{\mathcal N}));(T,Z)\mid(X,X_{\mathcal N})\right)=0. \end{align} We additionally assume that (ϕ(X),ϕ(XN))(\phi(X),\phi(X_{\mathcal N})) is a sufficient statistic of (X,XN)(X,X_{\mathcal N}) for (T,Z)(T,Z), i.e., \begin{align} (T,Z)\mathrel{\perp\perp}(X,X_{\mathcal N}) \mid(\phi(X),\phi(X_{\mathcal N})), \end{align} and hence \begin{align} I\left((X,X_{\mathcal N});(T,Z) \mid(\phi(X),\phi(X_{\mathcal N}))\right)=0. \end{align} Therefore, \begin{align} I((\phi(X),\phi(X_{\mathcal N}));(T,Z)) &=I((X,X_{\mathcal N}),(\phi(X),\phi(X_{\mathcal N}));(T,Z)) =I((X,X_{\mathcal N});(T,Z)). \end{align} This issue was also raised in [c].

(2) In Eqs. (42)-(44), let XX have density pp on Rd\mathbb{R}^d, and let Y=ϕ(X)Y=\phi(X) have density qq. Assume that the relevant differential entropies are well-defined. By the area formula, for almost every yy, \begin{align} q(y) = \sum_{x\in\phi^{-1}(y)} \frac{p(x)}{\left|\det D\phi(x)\right|}. \end{align} Therefore, for almost every xx, \begin{align} q(\phi(x)) \ge \frac{p(x)}{\left|\det D\phi(x)\right|}.
\end{align}

It follows that \begin{align} h(\phi(X)) &= \mathbb{E}\left[-\log q(\phi(X))\right] \le \mathbb{E}\left[ -\log p(X) + \log\left|\det D\phi(X)\right| \right] = h(X) + \mathbb{E}\left[ \log\left|\det D\phi(X)\right| \right]. \end{align} Since ϕ\phi is LL-Lipschitz, all singular values of Dϕ(x)D\phi(x) are at most LL almost everywhere. Hence, \begin{align} \left|\det D\phi(x)\right| \le L^d. \end{align} Consequently, \begin{align} h(\phi(X)) \le h(X)+d\log L . \end{align} This argument does not require ϕ\phi to be invertible. If ϕ\phi is many-to-one, the additional preimages only increase q(y)q(y), so the inequality still holds.

W2/Q1: Can the analysis cover the actual encoder and establish conditions under which the GW-only objective controls the omitted Frobenius term and the complete bound?

A: We appreciate the reviewer’s insightful observation. Because the Ricci-flow distance is updated as dij(s+1)=(1−ηκij(s))dij(s),d_{ij}^{(s+1)} =\left(1-\eta\kappa_{ij}^{(s)}\right)d_{ij}^{(s)}, and that, throughout SS iterations and η\eta steps, $\kappa_{\min} \le \kappa_{ij}^{(s)} \le 1. $

It follows that \begin{align} d_{ij}^{(S)}\le d_{ij}^{(0)} (1-\eta\kappa_{\min})^S=(1-\eta\kappa_{\min})^S=B_\phi. \end{align} Therefore, when the number of Ricci-flow iterations SS is finite, the step size is fixed, and the curvature admits a uniform lower bound, the resulting Ricci-flow distances are uniformly bounded. Furthermore, the entries of the two distance matrices satisfy 0≤Dijϕ≤B0\le D_{ij}^{\phi}\le B and 0≤Dijt≤10\le D_{ij}^{t}\le 1.

Let B=max⁡{Bϕ,1}B = \max \lbrace B_{\phi}, 1 \rbrace

Consequently,

\begin{align} \frac{1}{N^2} \left|D^{\phi}-D^t\right|_F^2 = \frac{1}{N^2} \sum_{i=1}^{N} \sum_{j=1}^{N} \left(D_{ij}^{\phi}-D_{ij}^{t}\right)^2 \le \frac{1}{N^2} \sum_{i=1}^{N} \sum_{j=1}^{N} B^2 = B^2. \end{align}

Thus, \begin{align} \frac{1}{N^2} \left|D^{\phi}-D^t\right|_F^2 \le \max \lbrace B_{\phi}, 1 \rbrace. \end{align} This establishes that the normalized Frobenius term is bounded under the finite-step lower-bounded-curvature assumptions.

Empirically, the normalized Frobenius term remains approximately 2.43×10−52.43\times10^{-5} to 2.2×10−52.2\times10^{-5} throughout training, whereas the total loss ranges from 0.20.2 to 0.40.4. It therefore accounts for approximately 0.005%0.005\% to 0.0001%0.0001\% of the total objective in this epoch. This observation suggests that the term makes a negligible numerical contribution under the current experimental setting.

W3: Redundant equations should be consolidated for clarity.

A: We appreciate your valuable suggestion. We will address this issue and revise the corresponding content in the updated manuscript.

Q3: Clarify the inconsistent exposure levels and effect definitions used in evaluation.

A: We appreciate the reviewer’s valuable comment. (1) Equation (77) contained a typographical error. The correct definition is εPEHETE=1N∑iN(τ^(xi)(1,1),(0,0)−τ(xi)(1,1),(0,0))2\varepsilon_{PEHE}^{TE}=\frac{1}{N}\sum_{i}^N(\hat \tau(x_i)^{(1,1),(0,0)}-\tau(x_i)^{(1,1),(0,0)})^2. Thus, both the estimated and ground-truth effects use the contrast between (1,1)(1,1) and ((0,0)((0,0) [a]. We have corrected the equation accordingly.

(2)We acknowledge that Appendix J previously and incorrectly referred to f(1,z)−f(0,z)f(1,z)-f(0,z), which follows the definition in [b], as the total effect. This quantity instead represents the direct treatment effect when the neighborhood exposure zz is held fixed. We will correct the definition in Appendix J and revised the corresponding experiments. The updated results and analysis of the real-world data are presented below:

The estimated total effect f(1,1)−f(0,0)f(1,1)-f(0,0) is −305.3-305.3, indicating that installing SCR/SNCR at a plant, together with adoption by its neighboring plants, reduces NOx emissions by an estimated 305.3 tons on average. This effect is larger in magnitude than the estimated direct effect of −221.6-221.6, suggesting an additional reduction of approximately 83.783.7 tons associated with neighborhood adoption. The result therefore provides evidence of beneficial spillover effects within the plant network.

[a] Learning individual treatment effects under heterogeneous interference in networks. TKDD 2024

[b] Doubly robust causal effect estimation under networked interference via targeted learning. ICML 2024

[c] Estimating Heterogeneous Treatment Effects: Mutual Information Bounds and Learning Algorithms. ICML 2023

Reviewer_cCwsOpenReview ↗

Follow-up after rebuttal

A deterministic encoder alone does not imply the mutual-information equality stated in the response. This step still requires an additional condition or correction, so I retain my rating of 2 for now.

We thank the reviewer and agree that a deterministic encoder alone is insufficient to establish the stated equality. We have therefore added the required sufficiency condition to the theorem: \begin{align} ( \phi(X),\phi( X_{\mathcal N})) \perp \perp (T,Z)\mid (X,X_{\mathcal N}) . \end{align}

Equivalently, \begin{align} I\left((X,X_{\mathcal N});(T,Z) \mid(\phi(X),\phi(X_{\mathcal N}))\right)=0. \end{align} Together with the determinism of ϕ\phi, the chain rule then yields the stated mutual-information equality.

Reviewer_F5jPOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

GEOMINI addresses causal effect estimation with network interference. It motivates representation balancing through a mutual-information PEHE bound and proposes maximizing a Gromov–Wasserstein discrepancy between representation and treatment geometries. Ricci-flow-refined graph distances are used for aggregation and exposure construction, and the model is trained with factual prediction loss and a geometric regularizer. Results on semi-synthetic BlogCatalog and Flickr data show promising direct-effect errors and some favorable spillover/total-effect results. However, the theoretical justification of the optimized surrogate, compatibility of the simulation with the identification assumptions, and consistency of the total-effect metric require substantial revision.

Critical strengths

S1. Relevant and difficult problem The work explicitly treats neighbors’ treatments and covariates as relevant to effect estimation, addressing an important limitation of independent-unit modeling.

S2. Interesting geometric construction Combining refined graph distances with distributional balancing is an interesting modeling direction. Several direct-effect results in Tables 1–2 are promising.

Critical weaknesses

W1. The optimized loss does not minimize the stated upper bound Equation (7) contains the positive representation-dependent term N⁻²‖Dϕ−Dᵗ‖²F, which is dropped because 1/N² is said to be small. The norm sums N² entries, so the normalized term is generally an average of order one. For distance matrices with off-diagonal distances 2 and 1 it equals 1−1/N. Maximizing only GW therefore need not decrease the displayed bound. Appendix I reports one empirical ablation, but that cannot justify the asymptotic argument or the claimed principled surrogate.

W2. The mutual-information argument has incompatible assumptions Theorem 4.2 assumes invertible ϕ, and Eq. (29) explicitly makes true mutual information invariant to that transformation; under this assumption it cannot be reduced through representation learning. If the implemented map is non-invertible, the proof needs different conditions. In addition, Appendix D.3 invokes nonnegative conditional entropy while subsequently using continuous differential entropy and a Jacobian transformation; differential conditional entropy is not generally nonnegative. Equation (6) also defines a squared difference of squared distances, whereas Eq. (57) derives a squared difference of distances. These issues affect the core theory, beyond notation alone.

W3. Simulation does not clearly satisfy overlap Equation (69) assigns treatment deterministically by whether the propensity score exceeds its mean, rather than sampling treatment from that propensity. Conditional on the fixed simulation and covariates, this conflicts with the overlap assumption used for identification. Equation (70) also uses constant direct treatment coefficients, so it provides limited evidence about heterogeneous direct effects. The paper should distinguish extrapolation under this simulator from identification under its assumptions and test a stochastic, overlap-preserving assignment.

W4. Total-effect evaluation is internally inconsistent Appendix A defines total effect at (1,z) versus (0,0). Appendix F Eq. (77) compares a prediction at z=1 with a ground-truth contrast at zᵢ, while Eq. (78) uses z=1 on both sides. The real-data discussion calls f(1,z)−f(0,z) a total effect, although this is a direct contrast at fixed exposure. These may be manuscript errors, but the implemented estimand and corrected results must be made explicit before interpreting total-effect performance.

Questions for authors

Q1. Can the authors provide a valid bound for the exact implemented representation class and GW loss, retain all representation-dependent terms, and establish conditions under which the optimized surrogate controls the desired dependence?

Q2. Does the implementation use the deterministic assignment of Eq. (69), and how do results change under stochastic treatment assignment with adequate overlap and heterogeneous effects?

Q3. Which exposure value is used on each side of the total-effect error in code? Please reconcile Eqs. (77)–(78), Appendix A, and the power-plant effect definitions, and identify any tables that must change.

3 comments

W1/Q1: The complete bound of 1/N21/N^2 Frobenius term.

A: We appreciate the reviewer’s insightful observation. Because the Ricci-flow distance is updated as dij(s+1)=(1−ηκij(s))dij(s),d_{ij}^{(s+1)} =\left(1-\eta\kappa_{ij}^{(s)}\right)d_{ij}^{(s)}, and that, throughout SS iterations and η\eta steps, $\kappa_{\min} \le \kappa_{ij}^{(s)} \le 1. $

It follows that \begin{align} d_{ij}^{(S)}\le d_{ij}^{(0)} (1-\eta\kappa_{\min})^S=(1-\eta\kappa_{\min})^S=B_\phi. \end{align} Therefore, when the number of Ricci-flow iterations SS is finite, the step size is fixed, and the curvature admits a uniform lower bound, the resulting Ricci-flow distances are uniformly bounded. Furthermore, the entries of the two distance matrices satisfy 0≤Dijϕ≤B0\le D_{ij}^{\phi}\le B and 0≤Dijt≤10\le D_{ij}^{t}\le 1.

Let B=max⁡{Bϕ,1}B = \max \lbrace B_{\phi}, 1 \rbrace

Consequently,

\begin{align} \frac{1}{N^2} \left|D^{\phi}-D^t\right|_F^2 = \frac{1}{N^2} \sum_{i=1}^{N} \sum_{j=1}^{N} \left(D_{ij}^{\phi}-D_{ij}^{t}\right)^2 \le \frac{1}{N^2} \sum_{i=1}^{N} \sum_{j=1}^{N} B^2 = B^2. \end{align}

Thus, \begin{align} \frac{1}{N^2} \left|D^{\phi}-D^t\right|_F^2 \le \max \lbrace B_{\phi}, 1 \rbrace. \end{align} This establishes that the normalized Frobenius term is bounded under the finite-step lower-bounded-curvature assumptions.

Empirically, the normalized Frobenius term remains approximately 2.43×10−52.43\times10^{-5} to 2.2×10−52.2\times10^{-5} throughout training, whereas the total loss ranges from 0.20.2 to 0.40.4. It therefore accounts for approximately 0.005%0.005\% to 0.0001%0.0001\% of the total objective in this epoch. This observation suggests that the term makes a negligible numerical contribution under the current experimental setting.

W2: The theory incorrectly treats invertibility as reducing mutual information and conditional differential entropy as nonnegative.

A: We thank the reviewer for identifying this issue. We agree that the change-of-variables argument in Eqs. (29) and (42)–(44) unnecessarily relies on the invertibility and nonnegative of ϕ\phi and is not needed in our discrete setting.

(1) In Eq. (29), since (ϕ(X),ϕ(XN))(\phi(X),\phi(X_{\mathcal N})) is determined by (X,XN)(X,X_{\mathcal N}), we have \begin{align} (\phi(X),\phi(X_{\mathcal N})) \mathrel{\perp\perp} (T,Z)\mid (X,X_{\mathcal N}), \end{align} and hence \begin{align} I\left((\phi(X),\phi(X_{\mathcal N}));(T,Z)\mid(X,X_{\mathcal N})\right)=0. \end{align} We additionally assume that (ϕ(X),ϕ(XN))(\phi(X),\phi(X_{\mathcal N})) is a sufficient statistic of (X,XN)(X,X_{\mathcal N}) for (T,Z)(T,Z), i.e., \begin{align} (T,Z)\mathrel{\perp\perp}(X,X_{\mathcal N}) \mid(\phi(X),\phi(X_{\mathcal N})), \end{align} and hence \begin{align} I\left((X,X_{\mathcal N});(T,Z) \mid(\phi(X),\phi(X_{\mathcal N}))\right)=0. \end{align} Therefore, \begin{align} I((\phi(X),\phi(X_{\mathcal N}));(T,Z)) &=I((X,X_{\mathcal N}),(\phi(X),\phi(X_{\mathcal N}));(T,Z)) =I((X,X_{\mathcal N});(T,Z)). \end{align} This issue was also raised in [e].

(2) In Eqs. (42)-(44), let XX have density pp on Rd\mathbb{R}^d, and let Y=ϕ(X)Y=\phi(X) have density qq. Assume that the relevant differential entropies are well-defined. By the area formula, for almost every yy, \begin{align} q(y) = \sum_{x\in\phi^{-1}(y)} \frac{p(x)}{\left|\det D\phi(x)\right|}. \end{align} Therefore, for almost every xx, \begin{align} q(\phi(x)) \ge \frac{p(x)}{\left|\det D\phi(x)\right|}.
\end{align}

It follows that \begin{align} h(\phi(X)) &= \mathbb{E}\left[-\log q(\phi(X))\right] \le \mathbb{E}\left[ -\log p(X) + \log\left|\det D\phi(X)\right| \right] = h(X) + \mathbb{E}\left[ \log\left|\det D\phi(X)\right| \right]. \end{align} Since ϕ\phi is LL-Lipschitz, all singular values of Dϕ(x)D\phi(x) are at most LL almost everywhere. Hence, \begin{align} \left|\det D\phi(x)\right| \le L^d. \end{align} Consequently, \begin{align} h(\phi(X)) \le h(X)+d\log L . \end{align} This argument does not require ϕ\phi to be invertible. If ϕ\phi is many-to-one, the additional preimages only increase q(y)q(y), so the inequality still holds.

W3/Q2: Clarify Eq. (69)’s deterministic assignment and validate the method under overlap and heterogeneous effects.

A: We appreciate this constructive suggestion. Following [b, c, d], our primary simulation assigns treatment by comparing each unit’s propensity score with the sample mean: \begin{align} T_i=\mathbb{I}(ps_i>\bar{ps}),\qquad \bar{ps}=\frac{1}{N}\sum_{i=1}^{N}ps_i.
\end{align} We additionally report results under a stochastic assignment mechanism, \begin{align} T_i\sim\operatorname{Bernoulli}(ps_i), \end{align} where each unit is treated with probability equal to its propensity score. The expected treated proportion under this mechanism is psˉ\bar{ps}. The additional experiment evaluates the robustness of our findings to different treatment-assignment mechanisms. The results are as follow:

RRNetTNetGDMLDWRGeomini
Direct PEHE (in)0.5356 ± 0.02590.4569 ± 0.10810.3040 ± 0.04800.3930 ± 0.04610.2849 ± 0.0698
Spillover PEHE (in)0.2297 ± 0.00080.1931 ± 0.07510.2148 ± 0.00940.2709 ± 0.05510.2266 ± 0.0032
Total PEHE (in)0.7233 ± 0.02780.8033 ± 0.14580.8261 ± 0.04580.3927 ± 0.14090.4227 ± 0.0976
Direct PEHE (out)0.5448 ± 0.02620.4618 ± 0.11120.2742 ± 0.05970.3901 ± 0.04310.2857 ± 0.0758
Spillover PEHE (out)0.2363 ± 0.00060.1918 ± 0.06470.2232 ± 0.00940.2545 ± 0.03300.2239 ± 0.0071
Total PEHE (out)0.7176 ± 0.02760.8006 ± 0.14390.7652 ± 0.13620.3952 ± 0.14100.4219 ± 0.0961
Direct ATE (in)0.5028 ± 0.02700.4352 ± 0.10200.3756 ± 0.05450.5204 ± 0.07310.2689 ± 0.0638
Spillover ATE (in)0.1368 ± 0.00070.1080 ± 0.04180.1019 ± 0.01710.0393 ± 0.04080.0855 ± 0.0222
Total ATE (in)0.6954 ± 0.02870.7946 ± 0.14140.4775 ± 0.03750.4381 ± 0.08160.4210 ± 0.0974
Direct ATE (out)0.5134 ± 0.02730.4423 ± 0.10460.2926 ± 0.14060.5233 ± 0.07640.2698 ± 0.0700
Spillover ATE (out)0.1460 ± 0.00060.1039 ± 0.02950.1191 ± 0.01520.0163 ± 0.01210.0840 ± 0.0131
Total ATE (out)0.6884 ± 0.02890.7921 ± 0.13970.4117 ± 0.13650.4568 ± 0.05540.4202 ± 0.0959

The results show that our method maintains its advantage under varying degrees of overlap and heterogeneous treatment effects.

W4/Q3: Clarify the total-effect estimand used in the power-plant effect definitions.

A: We appreciate the reviewer’s valuable comment.

(1) Equation (77) contained a typographical error. The correct definition is εPEHETE=1N∑iN(τ^(xi)(1,1),(0,0)−τ(xi)(1,1),(0,0))2\varepsilon_{PEHE}^{TE}=\frac{1}{N}\sum_{i}^N(\hat \tau(x_i)^{(1,1),(0,0)}-\tau(x_i)^{(1,1),(0,0)})^2. Thus, both the estimated and ground-truth effects use the contrast between (1,1)(1,1) and ((0,0)((0,0) [a]. We have corrected the equation accordingly.

(2)We acknowledge that Appendix J previously and incorrectly referred to f(1,z)−f(0,z)f(1,z)-f(0,z), which follows the definition in [b], as the total effect. This quantity instead represents the direct treatment effect when the neighborhood exposure zz is held fixed. We will correct the definition in Appendix J and revised the corresponding experiments. The updated results and analysis of the real-world data are presented below:

The estimated total effect f(1,1)−f(0,0)f(1,1)-f(0,0) is −305.3-305.3, indicating that installing SCR/SNCR at a plant, together with adoption by its neighboring plants, reduces NOx emissions by an estimated 305.3 tons on average. This effect is larger in magnitude than the estimated direct effect of −221.6-221.6, suggesting an additional reduction of approximately 83.783.7 tons associated with neighborhood adoption. The result therefore provides evidence of beneficial spillover effects within the plant network.

[a] Learning individual treatment effects under heterogeneous interference in networks. TKDD 2024

[b] Doubly robust causal effect estimation under networked interference via targeted learning. ICML 2024

[c] Estimating Causal Effects on Networked Observational Data via Representation Learning. CIKM 2022

[d] Generalization bound for estimating causal effects from observational network data. CIKM 2023

[e] Estimating Heterogeneous Treatment Effects: Mutual Information Bounds and Learning Algorithms. ICML 2023

Dear Reviewer F5jP,

We sincerely appreciate your detailed feedback and positive evaluation of our work. If any issues remain unresolved, we would be happy to further clarify them or provide additional information.

If our responses have satisfactorily resolved your concerns, we would be grateful for your support during the Reviewer–AC discussion period.

Best regards,

The Authors

Reviewer_323VOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes GEOMINI, a method for causal effect estimation under network interference. It uses Ricci flow to refine separate covariate and treatment graph structures, then maximizes their Gromov–Wasserstein discrepancy to learn balanced representations and aggregate neighbor information. The authors validate the method on semi-synthetic BlogCatalog and Flickr datasets and present an observational power-plant case study.

Critical strengths
  1. The paper tackles an important problem: estimating causal effects when individuals’ outcomes depend on their neighbors’ treatments.

  2. Separating the aggregation of covariates from that of treatments is a useful modeling idea. Applying graph geometry to both is an interesting direction.

  3. The semi-synthetic experiments cover two real network datasets and several levels of interference. GEOMINI performs well across many of the reported settings.

Critical weaknesses
  1. Theorem 4.2 assumes the representation map is invertible, so it preserves the input’s mutual information with treatment. The method, however, uses an unconstrained MLP and claims to learn treatment-invariant representations; the theorem therefore does not directly justify the representation learned in practice.

  2. The central claim that maximizing GW discrepancy reduces mutual information depends on Theorem 4.3, but its derivation needs stronger justification. In particular, the paper does not establish that the KDE error terms remain controlled as the representation is learned, and the optimization then drops a distance-matrix term from the bound.

  3. The construction of the treatment geometry is unclear. The Ricci-flow update requires initial treatment-graph distances, but the paper says all edges start with length 1 and does not explain how treatment values enter those distances; this makes the core treatment-specific geometry difficult to reproduce and assess.

  4. The causal evaluation has limitations. The main benchmarks use simulated treatments and outcomes, while the power-plant example has no known ground truth; moreover, Appendix F’s total-effect PEHE compares a prediction at (1,1)(1,1) with a target at (1,zi)(1,z_i), which appears inconsistent and could affect the reported results.

  5. The related-work section does not sufficiently cover or position the paper against closely related work on causal effect estimation under network interference and representation balancing.

Questions for authors

See Weaknesses.

3 comments

W1: The theorem’s invertibility assumption does not justify the treatment-invariant representation learned in practice.

A: We thank the reviewer for identifying this issue. We agree that the change-of-variables argument in Eqs. (29) and (42)–(44) unnecessarily relies on the invertibility and nonnegative of ϕ\phi and is not needed in our discrete setting.

(1) In Eq. (29), since (ϕ(X),ϕ(XN))(\phi(X),\phi(X_{\mathcal N})) is determined by (X,XN)(X,X_{\mathcal N}), we have \begin{align} (\phi(X),\phi(X_{\mathcal N})) \mathrel{\perp\perp} (T,Z)\mid (X,X_{\mathcal N}), \end{align} and hence \begin{align} I\left((\phi(X),\phi(X_{\mathcal N}));(T,Z)\mid(X,X_{\mathcal N})\right)=0. \end{align} We additionally assume that (ϕ(X),ϕ(XN))(\phi(X),\phi(X_{\mathcal N})) is a sufficient statistic of (X,XN)(X,X_{\mathcal N}) for (T,Z)(T,Z), i.e., \begin{align} (T,Z)\mathrel{\perp\perp}(X,X_{\mathcal N}) \mid(\phi(X),\phi(X_{\mathcal N})), \end{align} and hence \begin{align} I\left((X,X_{\mathcal N});(T,Z) \mid(\phi(X),\phi(X_{\mathcal N}))\right)=0. \end{align} Therefore, \begin{align} I((\phi(X),\phi(X_{\mathcal N}));(T,Z)) &=I((X,X_{\mathcal N}),(\phi(X),\phi(X_{\mathcal N}));(T,Z)) =I((X,X_{\mathcal N});(T,Z)). \end{align} This issue was also raised in [c].

(2) In Eqs. (42)-(44), let XX have density pp on Rd\mathbb{R}^d, and let Y=ϕ(X)Y=\phi(X) have density qq. Assume that the relevant differential entropies are well-defined. By the area formula, for almost every yy, \begin{align} q(y) = \sum_{x\in\phi^{-1}(y)} \frac{p(x)}{\left|\det D\phi(x)\right|}. \end{align} Therefore, for almost every xx, \begin{align} q(\phi(x)) \ge \frac{p(x)}{\left|\det D\phi(x)\right|}.
\end{align}

It follows that \begin{align} h(\phi(X)) &= \mathbb{E}\left[-\log q(\phi(X))\right] \le \mathbb{E}\left[ -\log p(X) + \log\left|\det D\phi(X)\right| \right] = h(X) + \mathbb{E}\left[ \log\left|\det D\phi(X)\right| \right]. \end{align} Since ϕ\phi is LL-Lipschitz, all singular values of Dϕ(x)D\phi(x) are at most LL almost everywhere. Hence, \begin{align} \left|\det D\phi(x)\right| \le L^d. \end{align} Consequently, \begin{align} h(\phi(X)) \le h(X)+d\log L . \end{align} This argument does not require ϕ\phi to be invertible. If ϕ\phi is many-to-one, the additional preimages only increase q(y)q(y), so the inequality still holds.

W2: Boundedness and Negligibility of the Ricci-Flow Distance-Matrix Term.

A: We appreciate the reviewer’s insightful observation. Because the Ricci-flow distance is updated as dij(s+1)=(1−ηκij(s))dij(s),d_{ij}^{(s+1)} =\left(1-\eta\kappa_{ij}^{(s)}\right)d_{ij}^{(s)}, and that, throughout SS iterations and η\eta steps, $\kappa_{\min} \le \kappa_{ij}^{(s)} \le 1. $

It follows that \begin{align} d_{ij}^{(S)}\le d_{ij}^{(0)} (1-\eta\kappa_{\min})^S=(1-\eta\kappa_{\min})^S=B_\phi. \end{align} Therefore, when the number of Ricci-flow iterations SS is finite, the step size is fixed, and the curvature admits a uniform lower bound, the resulting Ricci-flow distances are uniformly bounded. Furthermore, the entries of the two distance matrices satisfy 0≤Dijϕ≤B0\le D_{ij}^{\phi}\le B and 0≤Dijt≤10\le D_{ij}^{t}\le 1.

Let B=max⁡{Bϕ,1}B = \max \lbrace B_{\phi}, 1 \rbrace

Consequently,

\begin{align} \frac{1}{N^2} \left|D^{\phi}-D^t\right|_F^2 = \frac{1}{N^2} \sum_{i=1}^{N} \sum_{j=1}^{N} \left(D_{ij}^{\phi}-D_{ij}^{t}\right)^2 \le \frac{1}{N^2} \sum_{i=1}^{N} \sum_{j=1}^{N} B^2 = B^2. \end{align}

Thus, \begin{align} \frac{1}{N^2} \left|D^{\phi}-D^t\right|_F^2 \le \max \lbrace B_{\phi}, 1 \rbrace. \end{align} This establishes that the normalized Frobenius term is bounded under the finite-step lower-bounded-curvature assumptions.

Empirically, the normalized Frobenius term remains approximately 2.43×10−52.43\times10^{-5} to 2.2×10−52.2\times10^{-5} throughout training, whereas the total loss ranges from 0.20.2 to 0.40.4. It therefore accounts for approximately 0.005%0.005\% to 0.0001%0.0001\% of the total objective in this epoch. This observation suggests that the term makes a negligible numerical contribution under the current experimental setting.

W3: The treatment geometry is insufficiently specified, making it difficult to reproduce or evaluate.

A: Thank you for pointing out this ambiguity. Although all edges are initialized with dij(t)(0)=1d_{ij}^{(t)}(0)=1, the treatment values enter the Ricci-flow update when computing W(μi(t),μj(t))W(\mu_i^{(t)},\mu_j^{(t)}). Specifically, the transport cost between neighboring nodes uu and vv is calculated from their treatment values using a Hamming-type distance, i.e., cuv(t)=I(tu≠tv)=∣tu−tv∣c_{uv}^{(t)}=\mathbb{I}(t_u\neq t_v)=|t_u-t_v| for binary treatments. Therefore, different local treatment configurations produce different Wasserstein distances and Ricci curvatures κij(t)\kappa_{ij}^{(t)}, which subsequently update dij(t)d_{ij}^{(t)} through Eq. (10). The converged distances are then used to construct A~ijt\widetilde A_{ij}^{t} in Eq. (13). We will clarify this treatment-dependent ground cost in Section 4.2.

W4: The causal evaluation of the total effect has limitations.

A: We appreciate the reviewer’s valuable comment.

(1) Equation (77) contained a typographical error. The correct definition is εPEHETE=1N∑iN(τ^(xi)(1,1),(0,0)−τ(xi)(1,1),(0,0))2\varepsilon_{PEHE}^{TE}=\frac{1}{N}\sum_{i}^N(\hat \tau(x_i)^{(1,1),(0,0)}-\tau(x_i)^{(1,1),(0,0)})^2. Thus, both the estimated and ground-truth effects use the contrast between (1,1)(1,1) and (0,0)(0,0) [a]. We have corrected the equation accordingly.

(2)We acknowledge that Appendix J previously and incorrectly referred to f(1,z)−f(0,z)f(1,z)-f(0,z), which follows the definition in [b], as the total effect. This quantity instead represents the direct treatment effect when the neighborhood exposure zz is held fixed. We will correct the definition in Appendix J and revised the corresponding experiments. The updated results and analysis of the real-world data are presented below:

The estimated total effect f(1,1)−f(0,0)f(1,1)-f(0,0) is −305.3-305.3, indicating that installing SCR/SNCR at a plant, together with adoption by its neighboring plants, reduces NOx emissions by an estimated 305.3 tons on average. This effect is larger in magnitude than the estimated direct effect of −221.6-221.6, suggesting an additional reduction of approximately 83.783.7 tons associated with neighborhood adoption. The result therefore provides evidence of beneficial spillover effects within the plant network.

W5: Clarify GEOMINI’s distinction from existing network-interference and representation-balancing methods.

A: Thank you for this suggestion. We will further clarify that existing network representation-balancing methods generally aggregate covariates and treatments using the same network structure or aggregation mechanism. In contrast, GEOMINI explicitly models their different geometric structures, refines them separately through Ricci flow, and uses the Gromov–Wasserstein discrepancy to learn treatment-invariant representations.

[a] Learning individual treatment effects under heterogeneous interference in networks. TKDD 2024

[b] Doubly robust causal effect estimation under networked interference via targeted learning. ICML 2024

[c] Estimating Heterogeneous Treatment Effects: Mutual Information Bounds and Learning Algorithms. ICML 2023

Dear Reviewer 323V,

Thank you for your thoughtful review and encouraging assessment of our paper. Should you have any further questions or concerns, please feel free to let us know. We would be pleased to provide additional clarification.

If you believe that our responses have adequately addressed your concerns, we would sincerely appreciate your support for our paper during the Reviewer–AC discussion period.

Sincerely,

The Authors

Comments on the paper

1 comment

Dear AC,

We sincerely thank the reviewers for their constructive and insightful comments. We also highly appreciate their positive evaluations of our manuscript.

Specifically, Reviewer 323V: ''The paper tackles an important problem ...'' ''Separating the aggregation of covariates from that of treatments is a useful modeling idea ... '' ''GEOMINI performs well across many of the reported settings.''

Reviewer F5jP: ''... addressing an important limitation of independent-unit modeling.'' ''Combining refined graph distances with distributional balancing is an interesting modeling direction.'' ''Several direct-effect results in Tables 1--2 are promising.''

Reviewer cCws: ''Using separate geometries for representations and treatments is a well-motivated design ...' ''The experiments include several network causal estimators, with particularly competitive direct-effect results.''

The major concerns raised by the three reviewers largely overlap and center on the theoretical justification, the omitted distance-matrix term, the reproducibility of the treatment geometry, and the causal estimands and evaluation protocol. We summarize our responses and revisions below.

(1) Revising the theoretical analysis. In response to Reviewer 323V's W1, Reviewer F5jP's W2 and Reviewer cCWS's W1/Q2, we acknowledge that the original derivation unnecessarily relied on the invertibility of the representation map and the nonnegativity of differential conditional entropy, neither of which is required by our practical MLP encoder. We revise the analysis for a deterministic, potentially non-invertible representation using the data-processing relation and provide an entropy bound based on the area formula and the Lipschitz property of the encoder. The revised arguments no longer require an invertible representation and better align the theory with the implemented model.

(2) Accounting for the Frobenius term in the complete bound. In response to Reviewer 323V's W2, Reviewer F5jP's W1/Q1 and Reviewer cCWS's W2/Q1, we provide explicit conditions under which the Ricci-flow distances, and consequently the normalized Frobenius term, are uniformly bounded: a finite number of Ricci-flow iterations, a fixed step size, and a uniform lower bound on curvature. We also report its magnitude during training. The term remains approximately 2.2×10−52.2\times10^{-5}--2.43×10−52.43\times10^{-5}, compared with an overall objective of 0.20.2--0.40.4, showing that it is numerically negligible in the current experiments. We will state these conditions explicitly and clarify the relationship between the implemented GW surrogate and the complete bound.

(3) Clarifying the treatment-specific geometry. In response to Reviewer 323V's W3, Although treatment-graph edges are initialized with unit length, treatment information enters the Ricci-flow procedure through the ground transport cost cuv(t)=I(tu≠tv)c_{uv}^{(t)}=\mathbb{I}(t_u\neq t_v). Different local treatment configurations therefore induce different Wasserstein distances, Ricci curvatures, and updated treatment distances. We will add these details to make the method fully reproducible.

(4) Correcting and strengthening the causal evaluation. In response to Reviewer 323V's W4, Reviewer F5jP's W4/Q3 and Reviewer cCWS's Q3, 3e correct the total-effect estimand consistently to f(1,1)−f(0,0)f(1,1)-f(0,0) and revise the corresponding equations and real-world analysis. In the power-plant data, the corrected estimated total effect is −305.3-305.3 tons, compared with a direct effect of −221.6-221.6 tons, indicating an additional reduction associated with neighborhood adoption. In response to Reviewer F5jP's W3/Q2, we also add experiments using the overlap-preserving assignment Ti∼Bernoulli(psi)T_i\sim\mathrm{Bernoulli}(ps_i). The results show that GEOMINI remains competitive across direct, spillover, and total-effect estimation under the stochastic assignment mechanism.

(5) Clarifying the methodological contribution. In response to Reviewer 323V's W5, unlike existing network-balancing approaches that generally aggregate covariates and treatments using the same network structure, GEOMINI separately models and refines their geometric structures through Ricci flow and aligns them using the Gromov--Wasserstein discrepancy. We will emphasize this distinction and consolidate redundant equations to improve clarity.

Overall, the reviewers' comments have led us to correct the theoretical presentation and effect definitions, state the required assumptions more precisely, improve reproducibility, and add an overlap-preserving robustness experiment. All corresponding changes have been incorporated into the revised manuscript. These revisions substantially strengthen the rigor and clarity of the paper while preserving its central methodological contribution and main empirical findings.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.