When a treatment such as a lockdown, a drug, or a product feature is given to some units in a network, its effects spill over to their neighbors. This "network interference" breaks the textbook assumption that each unit's outcome depends only on its own treatment. Recent work found a geometric pattern in where estimation fails. Units sitting in tree-like bottlenecks, which have negative "Ricci curvature" (a local measure of how many redundant paths surround an edge), get much worse treatment-effect estimates than units in dense, triangle-rich regions. The proposed fix reweights the graph edges until the bottlenecks are smoothed out, a procedure called Ricci flow, and then trains the causal model on the smoothed graph. That fix rests on a confusion this note targets. In an interference problem the graph does two jobs. It is the physical channel through which spillovers actually travel, and it is also the scaffold a graph neural network uses to pool information about confounders. Smoothing a bottleneck may help the second job, but it rewrites the first. A bridge edge that carries all of the spillover between two communities is, causally, a bridge, and reweighting it changes the estimand.
We propose a continuous-time graph model that keeps the two roles apart. Its first component is a neural ordinary differential equation (a network that learns how a hidden state changes over time and is integrated with an ODE solver, so it can be queried at irregular times). This component propagates treatment exposure and outcomes over the observed graph, unchanged. A second, coupled representation channel evolves its own edge metric under a curvature-flattening flow and is used only for confounder pooling and treatment–control balancing. The central empirical bet has two parts. Applying Ricci flow to the spillover path, which is what existing curvature fixes do, should improve direct-effect estimates but bias spillover-effect estimates in exactly the bottleneck regions it targets. The decoupled model should get the direct-effect gains without that bias, and should do so in the irregularly sampled, time-varying settings where discrete-time models struggle.
Core Claim
Curvature-related error in network causal estimation has two different sources. One is representational: message passing loses information through bottlenecks, so confounders of bottleneck units are poorly summarized. The other is structural: spillovers through bottlenecks are rare, delayed, and single-path, so the data contain little information about them. Ricci flow can address the first. It cannot address the second, and applied to the spillover path it disguises the second by changing the effect being measured. A model that runs geometric smoothing only in a representation channel, while spillovers propagate on the true topology in continuous time, should therefore beat both a geometry-free continuous-time interference model and a geometry-preprocessed one on spillover effects, while matching the preprocessed model on direct effects.
Where This Sits
The motivating evidence comes from Farzam et al. In Curvature and Causal Inference in Network Data and From Geometry to Causality- Ricci Curvature and the Reliability of Causal Inference on Networks, they tie curvature to identifiability through entropy and anchor-regression arguments. They show that individual-effect error rises as curvature becomes more negative, and they build f-NetEst, which applies Ricci flow once, offline, before a graph-network estimator is trained. That procedure cuts EPEHE on BlogCatalog from 0.069 to 0.033 and reduces PEHE by up to 52%. We take their correlation as the phenomenon to explain. We differ in two ways. We never alter the graph on which spillovers are modeled, and we ask specifically whether their gain extends to peer effects, which their evaluation does not isolate.
The method for the representation channel comes from Graph Neural Ricci Flow: Evolving Feature from a Curvature Perspective. That paper evolves edge weights as functions of node features in continuous depth, proves that curvatures concentrate toward zero with bounded Dirichlet energy, and supplies EdgeNet, a linear-time curvature estimator. We reuse Attri-DRF and EdgeNet inside a causal model. Their paper has no treatments, no counterfactuals, and no time-varying graphs. Its finding that long integration horizons degrade separability also constrains our design, since in our setting the flow horizon must be a tuned hyperparameter rather than the physical time axis.
The continuous-time interference backbone comes from Causal Graph ODE: Continuous Treatment Effect Modeling in Multi-agent Dynamical Systems. CAG-ODE couples node and edge ODEs, fuses sparse policy events into continuous treatment signals, and balances representations adversarially. It reports roughly 3933 versus 4659 RMSE against a plain coupled graph ODE at 21 days on COVID-19 data. It uses no geometry, and it was tested only on 15- and 47-node systems, too small for curvature to vary meaningfully. The discrete-time counterpart, Modeling Interference for Treatment Effect Estimation in Network Dynamic Environment (DSPNET), uses graph convolutions plus recurrent units on fixed time bins, scales to Flickr-sized dynamic graphs, and also ignores geometry. R-ODE: Ricci Curvature Tells When You Will be Informed puts curvature into a graph ODE velocity field for cascade timing. It is purely predictive, and it lets curvature shape the propagation path itself, which is the design choice we argue is wrong for causal estimation.
Research Program
The model has three coupled states. For unit $i$, a representation state $h_i(t)$ pools confounders over a learned metric with edge weights $\omega_{ij}(t)$. An outcome state $y_i(t)$ is driven by the unit's own treatment $a_i(t)$ and by an exposure term computed on the observed adjacency $A_{ij}(t)$:
$$\frac{dh_i}{dt} = F_\theta\Big(h_i, \sum_j \omega_{ij}\,\phi(h_i,h_j)\Big), \qquad \frac{d\omega_{ij}}{d\tau} = -\hat\kappa_{ij}\,\omega_{ij},$$
$$\frac{dy_i}{dt} = G_\psi\Big(y_i,\; h_i,\; a_i(t),\; \sum_j A_{ij}(t)\, e_\eta\big(a_j(t), y_j\big)\Big).$$
Here $\hat\kappa_{ij}$ is the EdgeNet curvature estimate. The variable $\tau$ is a separate flow-depth variable, integrated for a tuned horizon at each observation time, rather than physical time $t$. The functions $\phi$ and $e_\eta$ are learned message and exposure functions, and $F_\theta, G_\psi$ are learned vector fields. The adversarial treatment-balancing loss of CAG-ODE is applied to $h$ only. The structural point is that $\omega$ never enters the outcome equation. Spillovers travel on $A$, which can change over time when edges are observed to appear or disappear, but which is never rewired.
The first part of the program is an analysis of where curvature error comes from. Using semi-synthetic graphs, we will generate outcomes from a known continuous-time diffusion of treatment effects over the true graph. We will then decompose each estimator's error into direct-effect and peer-effect parts, stratified by curvature bin, with node degree and local exposure count held fixed by matching. If the curvature–error relation disappears once degree and exposure counts are controlled, the geometric story reduces to a sample-size story, and we will report that.
The second part is the model itself and its ablations. Each ablation moves the Ricci flow to a different place: nowhere, onto both channels as static preprocessing in the f-NetEst style, onto both channels as a dynamic flow, or onto the representation channel only.
The third part is a secondary test of the "density penalty" tension. Causal Inference under Networked Interference and Intervention Policy Enhancement bounds policy error as worsening with density, while Farzam et al. associate dense, positively curved regions with easier identification. Our decomposition predicts that both can hold at once. Dense regions should be easy for representation and hard for disentangling many overlapping exposures. We test this by sweeping edge density while holding the curvature distribution fixed, as far as the generator allows.
The benchmarks are as follows. BlogCatalog and Flickr follow DSPNET's dynamic-edge protocol, extended with irregular observation times drawn from a point process, plus a spillover generator whose effect sizes differ between bridge and within-community edges. Cora and Chameleon, used by Farzam et al., provide contrasting curvature profiles. The PK-PD tumor simulation and the COVID-19 data from CAG-ODE are used as sanity checks. COVID-19 has no ground-truth counterfactuals, so it measures only factual long-horizon RMSE.
Baselines
The naive reference is a T-Learner given mean neighbor treatment as an extra covariate. Causal Inference under Networked Interference and Intervention Policy Enhancement reports that such non-graph estimators fail badly on Wave1, where its 1-GNN improves PEHE by 76.7%. We include this baseline so readers can see the size of the problem, not because we expect it to compete.
The strongest static-geometry competitor is f-NetEst from From Geometry to Causality- Ricci Curvature and the Reliability of Causal Inference on Networks, with its reported BlogCatalog EPEHE of 0.033. We will first reproduce that number in its original static setting before moving to our dynamic simulations.
The strongest dynamic discrete-time competitor is DSPNET, which reports $\sqrt{\epsilon_{PEHE}}=1.497$ on Flickr against 2.589 for DNDC. We likewise reproduce it in its own setting first.
The strongest continuous-time competitor is CAG-ODE. Its only reported number is the roughly 3933 RMSE at 21 days on COVID-19. We found no reported result on Lacuna for CAG-ODE on BlogCatalog or Flickr, so we will run it ourselves.
The key ablation is "CAG-ODE plus f-NetEst preprocessing": the same backbone, with the Ricci-flattened graph used for both representation and spillover. This isolates the decoupling. A second ablation applies the dynamic flow to both channels.
All methods share training splits, observation patterns, and a fixed tuning budget of the same number of hyperparameter trials per method on a validation set. Every ODE model uses the same solver and tolerance. We will report wall-clock time, because the stiffness problems noted in the GNRF paper may penalize our model.
Beating the baselines means the following. Peer-effect error in the most negative curvature quintile must fall significantly below the best baseline, across seeds and with degree held fixed. Overall direct-effect PEHE must stay within the variance of the both-channel flow variant.
Why not just…?
Why not just apply f-NetEst's static Ricci flow to CAG-ODE's graph? This is the most serious threat, and it is the main ablation. If static preprocessing matches our model on peer effects in bottleneck regions, then the decoupling claim is false and the contribution reduces to engineering. Our prediction is that preprocessing improves direct effects but pushes peer-effect estimates for bridge-adjacent units toward the community average, because it raises bridge edge weights. The stratified peer-effect error settles the question directly.
Why not just add curvature or degree as a node feature, or weight the balancing loss by curvature? This is cheap, and it might capture most of the representational benefit. We will include both as variants of CAG-ODE. If curvature-as-feature matches the representation-channel flow, we will adopt that simpler version, and the paper's claim will narrow to "decouple the spillover path, and annotate geometry rather than flowing it." The degree-as-feature variant also tests whether curvature adds anything beyond degree.
Why not just let attention learn edge weights? Attention can learn to up-weight bottleneck messages without any geometry. However, if it acts on the spillover path, it can learn to mismodel spillovers in exactly the way Ricci flow does, and if it acts only on the representation path, it becomes a learned version of our second channel. We therefore test attention in both placements. If attention restricted to the representation channel matches Attri-DRF, the evidence favors decoupling over geometry, and we will report it that way.
Why not just bin time and use DSPNET? DSPNET is a real competitor whenever observations are regular. We vary the irregularity of observation times from uniform to highly bursty. The continuous-time advantage should appear only as irregularity grows. If it does not appear, continuous time is not needed for this problem.
Decisive Checks
Three results decide the proposal.
First, in the matched-degree curvature analysis on BlogCatalog and Flickr, does f-NetEst-style preprocessing measurably increase peer-effect error in the most negative curvature quintile while lowering direct-effect error? If it does not, the premise fails, and we will report the simpler, purely representational account instead.
Second, does the decoupled model reduce bottleneck peer-effect error relative to CAG-ODE, CAG-ODE with preprocessing, and curvature-as-feature, while keeping direct-effect PEHE near that of the best flow variant? We also check that it matches CAG-ODE's factual RMSE on PK-PD and COVID-19, so that the geometry costs nothing where curvature is uninformative.
Third, under the density sweep, do direct-effect error and peer-effect error move in opposite directions? That would reconcile the density penalty with the curvature-robustness result, and would show the two were never measuring the same quantity.
References
- Curvature and Causal Inference in Network Data
- From Geometry to Causality- Ricci Curvature and the Reliability of Causal Inference on Networks
- Graph Neural Ricci Flow: Evolving Feature from a Curvature Perspective
- Causal Graph ODE: Continuous Treatment Effect Modeling in Multi-agent Dynamical Systems
- Modeling Interference for Treatment Effect Estimation in Network Dynamic Environment
- R-ODE: Ricci Curvature Tells When You Will be Informed
- Causal Inference under Networked Interference and Intervention Policy Enhancement