Graph neural networks (GNNs) classify a node by repeatedly averaging information from its neighbors. That averaging also distorts how confident the network is. Shallow GNNs on clean citation graphs tend to be underconfident. Deep GNNs over-smooth: node representations collapse toward one another, accuracy falls, and confidence climbs. When neighbors usually belong to different classes, a property called heterophily, averaging pushes nodes across decision boundaries. Current repair tools rescale confidence after training using a per-node "temperature", a divisor applied to the network's output scores before the softmax. Each of them predicts that temperature from a fixed graph snapshot and assumes that neighbors share confidence levels. That fails on heterophilous graphs. It also fails on graphs that change over time, such as Bitcoin transaction streams, where a model fitted on early months is applied to later months under new structure and no fresh labels.
We propose to predict each node's temperature from two cheap measurements, both recomputed directly from whatever graph exists at prediction time. The first is how often the node's k-hop neighbors agree with its predicted label according to the GNN. The second is how often they agree according to a feature-only classifier that never passes messages. When the GNN says a neighborhood agrees but the feature-only view says it does not, the consensus was produced by averaging rather than by the data. These are the over-smoothed, falsely confident nodes. When the GNN sees little 1-hop agreement but the feature-only view sees strong 2-hop agreement, the node sits in a heterophilous pattern the GNN partly exploits, and it is likely underconfident. Our central bet is that a calibrator built on these local, label-free signals transfers forward in time on a drifting graph without refitting, where calibrators tied to a fixed snapshot do not.
Core Claim
Much of the miscalibration that depth, heterophily, and temporal drift induce in GNNs is predictable from a node's two-view agreement profile. For node $i$ and hop $k = 1,\dots,K$, let $a_i^{(k)}$ be the fraction of $k$-hop neighbors whose GNN-predicted label matches $i$'s GNN-predicted label. Let $b_i^{(k)}$ be the same fraction computed with the feature-only classifier's labels. Let $d_i$ indicate whether the GNN and the feature-only classifier disagree on node $i$ itself. The calibrated prediction is
$$ \hat p_i = \mathrm{softmax}\!\left(\frac{z_i}{T_i}\right), \qquad T_i = \mathrm{softplus}\big(g_\theta(a_i^{(1:K)}, b_i^{(1:K)}, d_i, \log \deg_i)\big), $$
where $z_i$ is the frozen GNN's logit vector, $\deg_i$ is the node's degree, and $g_\theta$ is a small network fitted once on labeled validation nodes. Every input to $g_\theta$ comes from sparse multiplications of predicted one-hot label matrices by the current adjacency matrix. Nothing in it is spectral or tied to a fixed Laplacian, and it can be recomputed at each time step without labels.
The claim makes three predictions that can each fail on its own. On static heterophilous graphs, the profile should calibrate better than calibrators that assume confidence homophily. Across GNN depth, the gap $a_i^{(k)} - b_i^{(k)}$ should track the rise in overconfidence. On a drifting graph, a calibrator fitted on early time steps should hold its calibration error after a structural shock better than one recomputed from snapshot structure alone.
Relation to Prior Work
Three post-hoc calibrators motivate this design and set the bar it must clear. CaGCN trains an auxiliary GNN to output temperatures, assuming neighbors share confidence, and its authors flag heterophilous graphs as open. Ratio-Binned Scaling bins nodes by their predicted same-class neighbor ratio and fits one temperature per bin. That is essentially our $a_i^{(1)}$ alone, and its authors also leave heterophily untested. WATS feeds heat-kernel wavelet descriptors of node degree to a small network. It reports up to 41.2% lower ECE than baselines and shows wavelets beat degree or centrality descriptors, but only on static graphs. Our proposal differs from all three in one way. It adds a second labeling, the feature-only classifier's, that is immune to over-smoothing, and it uses multi-hop agreement rather than 1-hop agreement or degree structure. That second view is what can separate genuine neighborhood consensus from consensus manufactured by averaging. A single-view ratio cannot draw that line, because over-smoothing raises GNN agreement everywhere.
Two papers supply the mechanism. Two Sides of the Same Coin shows that low local homophily drives representations toward the opposing class's boundary. It also shows that over-smoothing behaves like "pseudo-heterophily": misclassified nodes become misleading neighbors and the error cascades. Our agreement gap is a direct, label-free detector of that cascade. Moderate Message Passing Improves Calibration shows that shallow propagation causes underconfidence and deep propagation causes overconfidence. Its fix, AU-LS, is a training-time label smoothing tested only on Cora, Citeseer, and Pubmed. We take its depth axis as an experimental variable and ask whether a post-hoc, structure-derived signal recovers the same correction without retraining. Following Overcoming the Homophily Assumption, we include 2- and 3-hop agreement because same-class neighbors often reappear beyond the first hop under heterophily.
The dynamic setting is where the gap is clearest. EvolveGCN, Skip-GCN as benchmarked in Anti-Money Laundering in Bitcoin, and THeGCN report only accuracy or $F_1$. None reports whether their probabilities stay calibrated over time, including across the Elliptic darknet shutdown at step 43, where predictive performance collapses. BaCa does test-time calibration with dynamic memory banks, but for graph-level out-of-distribution detection. It does not address node-level classification or heterophily. Adapting Graph Neural Networks to Evolving Topologies supplies the constraint that precomputed structural operators go stale when edges change. We design around that constraint rather than approximating it.
Research Program
The work has three parts.
The first is a measurement study the literature lacks: how calibration error evolves across time steps for dynamic GNNs on Elliptic and the PEMS traffic graphs, crossed with depth from 2 to 16 layers. We expect that overconfidence after step 43 will be concentrated in nodes with a large GNN-versus-feature agreement gap. If it is instead spread uniformly, the core claim is in trouble at the outset.
The second is the calibrator itself. It needs one feature-only classifier per backbone. On Elliptic this is naturally the Random Forest that already outperforms GNNs there. On the citation and web graphs it is a logistic-regression or MLP probe trained on the same labeled nodes. On top of that sits the small temperature network above, fitted on validation nodes from the training time window.
The third is a deployment protocol for drifting graphs. The calibrator is fitted once on Elliptic steps 1–34 and then applied, without labels, to steps 35–49. Its inputs are recomputed from each new snapshot. This protocol is compared against refitting strategies that consume delayed labels.
Two side questions are cheap to answer along the way. Does temporal message passing in THeGCN with negative attention weights slow the growth of the agreement gap with depth? Is the gap a better over-smoothing diagnostic than the Dirichlet-energy measures questioned in Are we measuring oversmoothing correctly?? For the second, we report correlation with actual calibration error and accuracy loss per node.
Baselines
The naive reference point is the uncalibrated backbone plus single-scalar temperature scaling fitted on the same validation nodes. It costs almost nothing, and any extra machinery must beat it. On Cora the uncalibrated GCN's ECE is reported as 23.51% in On Calibration of GNNs and 0.1347 for a 2-layer GCN in Moderate Message Passing. The two papers use different protocols, so we recompute it under ours rather than borrowing either.
The strongest static competitors are RBS, CaGCN, and WATS. Each is the most direct existing answer to "condition the temperature on local structure." RBS reports 3.90% ECE on Cora (with 3.10% when given ground-truth neighbor ratios) in On Calibration of GNNs. CaGCN is reported at 0.0401 in Moderate Message Passing. AU-LS, a training-time competitor, reports 0.0356 on Cora with a 2-layer GCN in the same paper. For WATS the file gives only the relative "up to 41.2%" figure; no reported exact ECE for WATS on heterophilous datasets was found on Lacuna. On Elliptic, no reported calibration result was found on Lacuna for any method. The accuracy references are EvolveGCN's illicit $F_1$ of 0.720 and Random Forest's 0.788 from Anti-Money Laundering in Bitcoin. Our calibrator does not change argmax predictions, so it must leave these unchanged by construction.
All post-hoc methods will share frozen backbones, identical validation splits, and the same hyperparameter search budget, with ten seeds each. For the dynamic protocol, every static calibrator gets the same treatment as ours: fitted on steps 1–34, with structural inputs recomputed per snapshot where the method allows. WATS is recomputed from each snapshot's degree wavelets, so it is not handicapped by stale structure. RBS and CaGCN are recomputed from each snapshot's predicted labels.
Two dynamic baselines consume delayed labels. One is rolling-window temperature scaling refit at each step using labels from the previous $m$ steps. The other is rolling-window RBS refit the same way. They represent what a practitioner would actually do.
Three ablations isolate the key ingredient. The first drops the feature-only view and uses $a^{(1:K)}$ alone. The second keeps both views but only at $k = 1$, which tests whether multi-hop matters. The third replaces the profile with degree, betweenness, and clustering coefficient, which tests whether we merely rediscover what WATS showed was too coarse.
Beating the competition means three things. On heterophilous datasets, ECE must be significantly below the best of RBS, CaGCN, and WATS. On Elliptic steps 43–49, class-wise ECE for the illicit class must be below every no-label baseline. And the first ablation must lose most of the gain.
Why not just…?
Why not just refit temperature scaling on a rolling window of recent labels? This is the most serious threat. If labels arrive with short delay, a global scalar refit every step may track drift well enough. The right response is to measure that in place of theory. We will vary the label delay $m$ and plot post-shock ECE for rolling temperature scaling against our zero-label calibrator. If rolling scaling with a delay of one or two steps matches us, the transfer claim survives only for long-delay regimes such as anti-money-laundering, where confirmations take months, and we will say so. A global scalar also cannot redistribute confidence between over-smoothed and heterophilous nodes in opposite directions. That is visible in the static results, where RBS beat temperature scaling in 11 of 15 settings in On Calibration of GNNs.
Why not just recompute RBS or WATS on each snapshot? Both can be recomputed cheaply. Honestly, for Elliptic-sized snapshots, sparse Chebyshev filtering is not prohibitive, so cost is not our argument. Our argument is that both condition only on GNN-view structure. Over-smoothing inflates GNN-view agreement everywhere, so a single-view calibrator should under-correct exactly the falsely confident nodes. The first ablation (our method without the feature-only view, which is nearly a continuous, multi-hop RBS) settles this directly. If it matches the full method, the second view is unnecessary and the contribution shrinks to multi-hop binning.
Why not just use the Random Forest with temperature scaling and skip the graph? On Elliptic, the Random Forest is more accurate than any GNN (illicit $F_1$ 0.788 against 0.720 for EvolveGCN). Its probabilities are untouched by message passing, so it may also be better calibrated. We will report its post-shock calibration with scalar and isotonic scaling. On Elliptic, its success depends on 72 hand-aggregated neighborhood features. On Chameleon and Cornell, structure-aware models beat feature-only baselines by 6% and 3% in Two Sides of the Same Coin. If the calibrated forest wins on Elliptic, the honest conclusion is that the feature-only view should be used for prediction there, with our profile as a disagreement detector. We would report that outcome.
Why not just keep GNNs at two layers? Shallow GCNs are systematically underconfident according to Moderate Message Passing, and heterophilous graphs need multi-hop context. Shallowness trades one calibration error for another rather than removing it. Our depth sweep reports the calibration of the best-calibrated shallow model alongside deeper calibrated ones, so a reader can see whether depth plus our calibrator ever beats simply staying shallow.
Proposed Experiments
The static half covers Cora, Citeseer, Pubmed, and CoraFull for the homophilous regime and Chameleon, Squirrel, Cornell, and Texas for the heterophilous regime. Backbones are GCN at depths 2, 4, 8, and 16, plus GGCN, so that we also test a model built to resist over-smoothing. Metrics are ECE, class-wise ECE, negative log-likelihood, and ECE stratified by bins of true local homophily. The stratified view is where confidence-homophily methods should fail and ours should not.
The dynamic half uses Elliptic, with the 34/15 temporal split from the anti-money-laundering benchmark, and PEMSLA, where THeGCN reports 55.82% accuracy. Backbones are EvolveGCN, Skip-GCN, and THeGCN. Because Elliptic is heavily imbalanced, the main metric is illicit-class reliability, measured as class-wise ECE and Brier score on illicit-labeled nodes, reported per time step.
Three checks are decisive. First, before building anything, the per-node agreement gap $a_i^{(k)} - b_i^{(k)}$ must predict per-node overconfidence, measured as confidence minus correctness, with rank correlation clearly above that of degree or 1-hop GNN agreement. This must hold on the deep backbones and on Elliptic steps 43–49. Second, the no-feature-view ablation must lose most of the improvement on heterophilous graphs and deep backbones. Third, the zero-label calibrator fitted on steps 1–34 must keep post-shock illicit-class ECE below snapshot-recomputed RBS and WATS, and the label delay at which rolling temperature scaling catches up must be reported. If the first check fails, the project reduces to a measurement study of calibration drift in dynamic GNNs. That study is still unreported in the literature.
References
- Be Confident! Towards Trustworthy Graph Neural Networks via Confidence Calibration
- On Calibration of Graph Neural Networks for Node Classification
- WATS: Wavelet-Aware Temperature Scaling for Reliable Graph Neural Networks
- Two Sides of the Same Coin: Heterophily and Oversmoothing in Graph Convolutional Neural Networks
- Moderate Message Passing Improves Calibration: A Universal Way to Mitigate Confidence Bias in Graph Neural Networks
- EvolveGCN: Evolving Graph Convolutional Networks for Dynamic Graphs
- Anti-Money Laundering in Bitcoin: Experimenting with Graph Convolutional Networks for Financial Forensics
- THeGCN: Temporal Heterophilic Graph Convolutional Network
- Graph Out-of-Distribution Detection via Test-Time Calibration with Dual Dynamic Dictionaries
- Are we measuring oversmoothing in graph neural networks correctly?