Many treatments in networks are assigned by position, not by neighborhood. A hospital's referral hubs get the new protocol first. A company promotes the employees who bridge departments. Students who sit at the edge of their friendship circle are the ones a school counsellor targets. In each case, a node's role in the network (hub, bridge, periphery) drives who gets treated, and that same role often affects the outcome. Two nodes can share a role while sitting far apart, or even in separate networks. Current network-aware methods for causal inference mostly assume confounding spreads through proximity: connected people share hidden traits. They build their adjustment from neighborhood averages or community memberships, and a hub in one community looks nothing like a hub in another.
We propose to estimate propensity scores and form matches using structural signatures: vectors that describe the shape of a node's surroundings and do not depend on where the node sits. A propensity score is the probability of being treated given what we observe; matching treated and untreated units with similar scores removes measured confounding. Our signatures come from diffusion wavelets, which summarize how heat spreads outward from each node. The second ingredient is the key one. We do not force treated and control groups to agree on the whole signature, which would erase role differences that legitimately predict the outcome. Instead we balance only the few directions of the signature along which the two groups actually differ. Because signatures need no shared coordinates, the method can match a treated student in one school to a control student in another school by role.
Core Claim
When treatment depends on multi-hop structural role, three standard adjustments leave bias that matching on a diffusion-wavelet signature removes:
- logistic propensity scores with hand-computed centralities,
- graph neural network deconfounders that average over neighbors,
- community random effects.
Balancing that signature only in its worst-imbalanced low-dimensional subspace keeps the estimate unbiased without discarding outcome-relevant structure. When confounding is purely degree-driven or purely homophilous, the method should match the simpler adjustments rather than beat them. Showing where it wins and where it merely ties is part of the claim.
Relation to Prior Work
The proposal reuses its structural machinery from two papers.
- GraphWave embeds each node by sampling the characteristic function of its heat-wavelet coefficients. It does this in time linear in the number of edges, and it guarantees that nodes with isomorphic neighborhoods get identical embeddings. That gives us a covariate-like vector that is comparable across disconnected graphs.
- struc2vec aligns hop-by-hop degree sequences with dynamic time warping, a method for comparing sequences of unequal length. It showed that role similarity carries across disconnected components. We use it as an alternative signature, not the main one, because it took 260 seconds on a 100-node graph.
Neither paper does causal inference. Their benchmarks score role recovery, not treatment-effect bias.
Three causal papers motivate the gap.
- Network Deconfounder learns hidden confounders with a graph convolutional network, which averages features over neighbors. It then penalizes the Wasserstein distance between treated and control representations across the entire latent space. Both choices are what we change: neighbor averaging cannot see role, and full-space balancing invites over-balancing.
- Bayesian generalized propensity scores under interference adds community random effects, which capture dense clusters but not roles that cut across them.
- Toulis et al. show that independent-unit logistic propensity models are badly misspecified when treatment is entangled with topology. They fix this by simulating network evolution with generative models (exponential random graph or latent space models), which scale poorly. We ask whether a non-parametric signature recovers most of that correction without simulation.
Our subspace step comes from Spiked-CFR. That work balances only a worst-case k-dimensional projection of high-dimensional text representations. It argues that full-dimensional Wasserstein balancing converges at rate $n^{-1/D}$ and destroys information that predicts the outcome. We apply the same logic to structural signatures, which have the same failure mode. Integral-probability-metric balancing and optimal balancing conditions supply the general idea of building balance into score estimation; neither touches network structure.
Finally, SITE handles scale through linear pre-aggregation of neighbor features. It assumes no unobserved confounding and is again neighborhood-based. We borrow its emphasis on CPU-scale computation.
Research Program
Step 1: compute signatures. For each node we compute a GraphWave signature $z_i$ at a small set of diffusion scales. Choosing the scale is a known sensitivity problem. We choose it at the design stage, without looking at outcomes, by maximizing how well $z$ predicts treatment out of sample. This respects the matching principle that the design is fixed before outcomes are seen.
Step 2: fit the propensity score. We fit $e(x_i, z_i)$, where $x_i$ are node attributes, using regularized logistic regression or gradient-boosted trees.
Step 3: match and refine. We form caliper matches on $e$. We then run projection pursuit on the matched sample. We find the $k$-dimensional orthonormal projection $W$ of $z$ that maximizes
$$ \mathcal{W}_1\big(W^\top z_{\text{treated}},\; W^\top z_{\text{matched control}}\big), $$
where $\mathcal{W}_1$ is the Wasserstein-1 distance between the projected treated and control distributions. We then re-match using Mahalanobis distance on $(e, W^\top z)$. We iterate until the worst-case projected imbalance falls below a permutation-calibrated threshold, or until $k$ reaches a small cap.
Directions of $z$ along which the groups already agree are never forced together. This is the safeguard against over-balancing.
Disconnected networks. Matching uses only $(x, e, z)$, never graph coordinates. So it applies directly across disconnected networks, such as separate schools. In that setting we compute signatures per network and pool the units.
Theory. The theoretical piece is modest and specific. If treatment depends on structure only through a function of the $K$-hop neighborhood up to isomorphism, then the wavelet signature at matching scales determines that function on isomorphism classes. This follows from GraphWave's equivalence guarantee. Matching on $(x, z)$ then inherits the usual ignorability argument, conditional on observed attributes and role. We will state where the guarantee breaks: the signature is only approximately injective, so two different neighborhoods can map to nearby signatures. We will measure how much bias that leaves.
Baselines
Naive reference. The unadjusted difference in means, and logistic propensity matching on node attributes $x$ alone. These fix the size of the bias to be removed.
Strongest simple competitor. Logistic propensity matching on $x$ plus hand-computed structural features: degree, PageRank, betweenness, clustering coefficient, eigenvector centrality, and two-hop degree statistics. It is the strongest competitor because many roles (hubs especially) are largely summarized by degree and centrality. It receives the same matching protocol, caliper, and tuning budget for its feature set as our method gets for diffusion scales.
Strongest learned competitor. The Network Deconfounder, run with its published objective. We also run a variant whose input features are augmented with degree. That gives the neighbor-averaging model a fair chance to see structure. On BlogCatalog and Flickr the file reports only that it beat CFR-Wass, CFR-MMD, TARNet, CEVAE, Causal Forest and BART. No reported numeric result was found on Lacuna for its error on BlogCatalog or Flickr, so we rerun it rather than quote it.
Other learned and structural competitors. - SITE on the larger graphs. - A heterophily-capable network from the CMGNN codebase, as a learned encoder that is not purely homophilous. - Community random effects in the style of Bayesian GPS.
Entangled treatments. The reference is Toulis et al.. They report average-treatment-effect RMSE of 57.05 for the classical misspecified propensity model and 3.41 for their network-evolution model on 100-unit simulations. We do not expect to beat 3.41 on data generated by their own model. Closing most of the gap from 57.05 without fitting a generative network model would count as success.
All learned models get the same hyperparameter-search budget, measured in GPU-hours, and the same train and validation splits. All matching methods share caliper width and matching ratio.
Ablations. Two ablations isolate the key ingredient. The first uses GraphWave signatures in a logistic propensity score with no subspace refinement. The second balances the full signature, Network-Deconfounder style, instead of the worst-case subspace.
What counts as a win. Beating the baselines means lower average-treatment-effect bias and lower PEHE (precision in estimating heterogeneous effects, the root-mean-square error of individual effect estimates). It must also be achieved at comparable or higher retained sample size, judged by bootstrap confidence intervals that exclude zero difference across seeds.
Why not just…?
Why not just add centrality measures to a logistic propensity model? This is the real threat. If roles in practice are mostly "how connected you are," it may be enough. GraphWave's evidence against hand-built features, such as Enron completeness of 0.067 versus 0.028 for RolX, concerns role discovery, not confounding adjustment, so it does not settle the question. Our benchmarks therefore include a regime where treatment depends only on degree, where we expect a tie. They also include a regime where it depends on multi-hop shape at fixed degree, where we expect centralities to leave bias. If centralities match our method in both regimes, we will report that hand-built features suffice and the signature adds nothing.
Why not just use a graph neural network deconfounder, maybe one built for heterophily? Neighbor-averaging encoders give structurally identical but distant nodes different representations unless their features happen to agree. The struc2vec and GraphWave synthetic results show this failure for homophily-based embeddings. A heterophily-aware network or degree-augmented inputs might partly fix this, which is exactly why both are run as baselines. Cross-network matching, as in Add Health schools, is a further test. A model trained per graph has no shared coordinate system, whereas our signatures do.
Why not just add community random effects? Communities are proximity clusters. A bridge in one cluster and a bridge in another get unrelated effects, so role confounding survives. Our planted-role benchmark makes role and community nearly independent, so this baseline should fail there. It should succeed in the homophily regime, where we expect it to beat us. We will report that loss too.
Why not just balance the whole signature? Full-space Wasserstein balancing is the ablation. Spiked-CFR's argument predicts that it erases outcome-predictive structure and raises PEHE even when average-effect bias looks fine. If it performs as well as subspace balancing, the projection step is unnecessary and we drop it.
Proposed Experiments
Benchmark 1: planted roles on real graphs. We follow GraphWave's construction, attaching small motifs (houses, cycles, stars) to the BlogCatalog and Flickr graphs, so ground-truth roles are known and defined independently of any signature method. Treatment depends on attributes and planted role, and outcomes depend on both. We vary three conditions:
- the strength of role confounding,
- whether roles are separable by degree alone,
- the mix of role-driven and community-driven assignment.
We also reproduce the original Network Deconfounder homophily setting to check we do not lose there.
Benchmark 2: Toulis entanglement simulation. We reproduce the Toulis et al. 100-unit entanglement simulation and compare RMSE against their 57.05 and 3.41.
Benchmark 3: Add Health. On Add Health, treatment is semi-synthetic and depends on friendship-network position within schools. Matching is restricted to across-school pairs, a setting that proximity-based methods cannot handle.
Scaling check. On Twitch and AMZ-P we check that signature computation plus matching stays within SITE's reported regime of 46 seconds on AMZ-P. Graph tooling comes from CogDL.
Metrics. We report:
- average-treatment-effect bias and RMSE,
- PEHE,
- standardized mean differences on the hidden role variable after matching (the role is unseen by all methods),
- the fraction of treated units retained under the caliper.
Decisive checks. First, a regime-by-method interaction. Our method should beat centrality-augmented propensity matching only when roles are not degree-separable, and beat the Network Deconfounder only when assignment is role-driven rather than community-driven. A uniform win would itself suggest a benchmark artifact. Second, the subspace ablation should lower PEHE relative to full balancing at equal bias. Third, across-school matching should give lower bias than within-school matching restricted by sample size. If any of these fails, the corresponding piece of the proposal is cut.
References
- Learning Structural Node Embeddings via Diffusion Wavelets
- struc2vec: Learning Node Representations from Structural Identity
- Learning Individual Causal Effects from Networked Observational Data
- Estimating Causal Effects under Network Interference with Bayesian Generalized Propensity Scores
- Propensity score methodology in the presence of network entanglement between treatments
- Spiked-CFR: Causal Representation Learning from LLMs via Wasserstein Projection Pursuit
- Covariate balancing using the integral probability metric for causal inference
- Optimal Covariate Balancing Conditions in Propensity Score Estimation
- Scalable individual treatment effect estimator for large graphs