- 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
-
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.
-
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
-
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.
-
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.
-
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
-
Can the analysis cover the actual encoder and specify a condition that controls the omitted matrix term during optimization?
-
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?
-
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
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 . Thus, both the estimated and ground-truth effects use the contrast between and [a]. We have corrected the equation accordingly.
(2)We acknowledge that Appendix J previously and incorrectly referred to , which follows the definition in [b], as the total effect. This quantity instead represents the direct treatment effect when the neighborhood exposure 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 is , 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 , suggesting an additional reduction of approximately 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
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 , the chain rule then yields the stated mutual-information equality.
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 and is not needed in our discrete setting.
(1) In Eq. (29), since (ϕ(X),ϕ(XN)) is determined by (X,XN), 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)) is a sufficient statistic of (X,XN) for (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 X have density p on Rd, and let Y=ϕ(X) have density q. Assume that the relevant differential entropies are well-defined. By the area formula, for almost every y, \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 x, \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 is L-Lipschitz, all singular values of Dϕ(x) are at most L 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 to be invertible. If ϕ is many-to-one, the additional preimages only increase 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), and that, throughout S iterations and η 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 S 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≤B and 0≤Dijt≤1.
Let B=max{Bϕ,1}
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 2.43×10−5 to 2.2×10−5 throughout training, whereas the total loss ranges from 0.2 to 0.4. It therefore accounts for approximately 0.005% to 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.