Graph neural networks learn about a node by repeatedly averaging it with its neighbors. That works when linked nodes tend to share a label, which is the homophily assumption. It fails when linked nodes tend to differ, which is called heterophily. It also fails when the averaging is repeated so often that every node ends up looking the same, a collapse called over-smoothing. On real systems that change over time, such as traffic sensors, transaction streams, or comment threads, the fraction of "agreeing" neighbors is not fixed. It drifts, and it can jump suddenly. So the right amount of averaging for a node changes from hour to hour. This note asks whether the curvature of the embedding space can serve as that per-node, per-moment amount of averaging. Curvature describes how fast space opens up with distance: flat space is ordinary Euclidean space, and negatively curved (hyperbolic) space gains room exponentially. The curvature would be learned and updated online, rather than fixed, as it is in most hyperbolic models, or routed only on topology, as in current dynamic curvature models.
The opening is a contradiction the literature has noticed but not resolved. In flat space, deep averaging squeezes representations together into a point. In hyperbolic space, H4G finds the opposite failure: embeddings drift toward the boundary of the Poincaré ball, and fine local distinctions are compressed into coarse "over-abstracted" summaries, which hurts heterophilic graphs most. If the two failures sit at opposite ends of a curvature axis, then a node's curvature is a natural control variable between them. If they don't, curvature adds nothing that a plain learned mixing weight cannot do. We design the work so that either answer is informative. The central test pits the curved model against a Euclidean gate that receives exactly the same control signal.
Core Claim
We claim the following. A temporal message-passing network can carry a per-node curvature $\kappa_i(t)$ set by a small controller. That controller reads two streaming, label-free quantities that Two Sides of the Same Coin identified as the joint root causes of heterophily failure and over-smoothing. The first is estimated local homophily, computed from current feature or predicted-label agreement. The second is relative degree. Such a network will avoid both flat-space collapse and hyperbolic boundary drift under drifting homophily. It will also recover faster after heterophilic shocks than either a fixed-curvature model or a Euclidean model whose per-node aggregation gate is driven by the same controller.
What is new is the coupling. Curvature is driven by homophily and degree, not by topology alone. It is updated continuously inside an end-to-end trained temporal backbone. And it is judged with a two-sided collapse diagnostic that measures both failure modes at once.
Concretely, the controller sets
$$\kappa_i(t) = -\,\mathrm{softplus}\big(g_\phi(\hat h_i(t),\, \bar r_i(t),\, \kappa_i(t^-))\big),$$
where $\hat h_i(t)$ is the estimated homophily of node $i$'s recent neighborhood, $\bar r_i(t)$ is its degree relative to its neighbors' degrees, $\kappa_i(t^-)$ is its previous curvature, and $g_\phi$ is a small multilayer perceptron. A sign-free variant also allows $\kappa \ge 0$, which is spherical geometry. Aggregation happens in node $i$'s tangent space at its own curvature, using signed attention weights in the style of THeGCN.
Relation to Prior Work
The closest work falls into three groups.
The first group has dynamic curvature but no homophily. CurvPrompt routes prompts over a frozen temporal backbone to hyperbolic, flat, and spherical experts using local topology. DyGMoCE routes inside a windowed dynamic graph transformer under a Gromov δ-hyperbolicity constraint. Both route on structure alone and assume homophilic aggregation. DyGMoCE also reports that its experts become redundant. We hypothesize that this redundancy happens because topology cannot tell a branching hierarchy from a heterophilic burst, and that a homophily signal can.
The second group has a curvature–smoothing link but is static. GNN+BEC learns Bakry-Émery curvature, a curvature that depends on how features diffuse and not only on graph shape. It uses that curvature to stop message passing early at fast-equilibrating nodes, partitioning curvature into layer cutoffs with a hand-set threshold. We take from it the idea that curvature predicts feature decay. We differ by using curvature as a continuous geometric control in time, not a discrete depth cutoff on a static graph. CUSP combines mixed-curvature product spaces with spectral filterbanks. Its finding that Riemannian models act as low-pass filters motivates our signed aggregation. Its ablation, where removing curvature costs about 4% on heterophilic data, is the main existing evidence that geometry adds something beyond filtering, but only on static graphs. Neural Sheaf Diffusion shows that a richer edge geometry can increase Dirichlet energy and so "un-smooth" features. It constrains our design: any gain we claim must exceed what a learned non-scalar diffusion already gives.
The third group handles temporal heterophily but stays Euclidean. THeGCN handles temporal edge heterophily with signed spatio-temporal attention entirely in flat space. It is our backbone template and our strongest competitor.
Two constraints frame the evaluation. Temporal Graph Benchmark shows that non-learning heuristics often beat temporal neural networks. Are we measuring oversmoothing correctly? warns that similarity metrics such as Dirichlet energy confuse useful refinement with harmful collapse. So our diagnostics are always paired with accuracy against held-out labels.
Research Program
The work proceeds in two stages, and the second depends on the first.
Stage one: a controlled measurement study of curvature and smoothing. We build synthetic dynamic graphs from planted partitions. In these graphs, the intra-class and inter-class edge probabilities follow scripted trajectories: slow drift, periodic inversion, and sudden shocks. Some components are tree-like and some are cyclic, so that topology-derived curvature and homophily disagree by construction. For networks of fixed curvature from −2 to +1, and from depth 2 to 64, we measure two quantities after each layer. The first is the tangent-space Dirichlet energy, which captures squeezing together. The second is the mean distance of embeddings from the manifold origin relative to the maximum, which captures H4G-style boundary drift. Both are measured alongside linear-probe accuracy.
The question is whether there is a curvature region where both measures stay healthy, and whether that region moves predictably with homophily. If it does not move, curvature is not a useful dial, and we report that negative result instead of building stage two.
Stage two: the online curvature controller. If stage one shows that the healthy region tracks homophily, we build the controller-driven temporal network described above. It uses THeGCN's sliding-window sampler, cosine time encoding, and signed attention, but runs in per-node κ-stereographic geometry. We add a fast polar-coordinate attention in the style of DyGMoCE to keep the overhead of operations on the curved space manageable. Homophily is estimated without test labels, from agreement between the model's own predicted labels or features over a decayed window. Training is end-to-end. There is no auxiliary distortion loss, which avoids the loss-weighting sensitivity CurvPrompt reports. The only regularizer is a smoothness penalty on $\kappa_i(t)$ over time.
Baselines
The naive reference points come from two families. A structure-ignoring multilayer perceptron answers whether message passing helps at all under heterophily; Two Sides of the Same Coin found that most GNNs lose to it. On temporal graph benchmark node-affinity tasks, the Moving Average and persistence heuristics are the do-nothing references. On tgbn-reddit, Temporal Graph Benchmark reports NDCG@10 of 0.559 for Moving Average against 0.315 for TGN. Any gain on these tasks must be stated relative to the heuristic, not to TGN.
The strongest existing method on temporal heterophily is THeGCN, which reports 55.82% accuracy on PEMSLA against 48.27% for APAN. It is the right competitor because it already has signed, temporally aware filtering, so any gain we show cannot come from merely allowing negative messages. We rerun it with the authors' sampler on our splits. It gets the same number of tuning trials and the same parameter budget, within 5%, as our model, and results are averaged over at least five seeds. Beating it means a mean improvement larger than the combined seed standard deviation on PEMSLA, and the same on at least two of PEMSBA, PEMSSD, and Reddit.
Among curvature-aware dynamic models, we compare against DyGMoCE. The file reports it as ranked first on almost all of its eight datasets, but no reported result on THeGCN's heterophilic benchmarks was found on Lacuna for DyGMoCE. We run it with its released router on our splits under the same tuning budget. We also include CurvPrompt with its backbone unfrozen, so that its frozen-backbone design is not an unfair handicap.
For the static curvature–smoothing link, GNN+BEC is applied snapshot by snapshot. The file reports a 13–16% gain over GCN on Texas and Wisconsin for this method, and we use those two datasets as a static sanity check that our reimplementation is faithful.
Three ablations isolate the key ingredient. In the matched Euclidean gate, the same controller output sets a per-node scalar that mixes self-information and neighbor-information in flat space, with identical inputs and parameter count. The second ablation uses a single learnable global curvature, in the style of Hypformer's learned layer curvatures. The third is a topology-only controller that is fed δ-hyperbolicity instead of homophily.
Why not just…?
Why not just use Euclidean signed attention with a learned per-node gate, as THeGCN effectively does? This is the real threat. A scalar gate driven by the same homophily signal might capture all of the benefit, and curvature would then be decoration. The matched Euclidean gate ablation settles this directly, because it receives the identical controller signal. We commit in advance to the following rule: if the curved model does not beat the matched gate by more than the seed variation on at least half of the temporal tasks, we conclude that curvature is unnecessary for drifting homophily. The only evidence in its favor so far is CUSP's static ablation. We expect curvature to matter specifically where tree-like structure and heterophily coexist, because exponential volume can keep hierarchy levels apart while the gate decides how much to mix. The synthetic stage tests exactly this.
Why not just run a heuristic like Moving Average or EdgeBank? On several temporal graph benchmark tasks these win outright, at a tiny fraction of the cost. We report them on every node-affinity task and do not claim progress where they win. Our target regime is abrupt heterophilic inversion, as in PEMS congestion flips. Persistence-style predictors are wrong by construction immediately after such a flip, and we measure recovery time after scripted shocks to show this.
Why not just add residual connections or PairNorm? Residual connections and PairNorm prevent global collapse, but they do not stop discordant neighbors from contaminating a node. GNN+BEC's 64-layer stress tests show deep residual models degrading on heterophilic graphs. We include residual plus PairNorm versions of THeGCN in the depth sweep. The two-sided diagnostic will show whether they keep energy healthy while accuracy still falls.
Why not just compute temporal curvature offline, as TRicci does? TRicci keeps 97.7% of ROC-AUC with 80% of edges pruned using temporal Forman-Ricci curvature. That curvature reads only graph shape, however, and is blind to whether neighbors agree. We test a variant in which the controller is fed TRicci curvature instead of estimated homophily. If it matches our model, the homophily signal is not what matters.
Proposed Experiments and Decisive Checks
The first decisive check is the synthetic stage. We test whether the curvature region that avoids both collapse modes shifts monotonically with planted homophily. We also test whether the learned $\kappa_i(t)$ follows the scripted homophily trajectory, measured as the correlation between the two series and the lag after shocks.
The second check is the matched-gate comparison on PEMSLA, PEMSBA, PEMSSD, Reddit, and Brain, reporting accuracy and accuracy recovery time after the largest homophily drops.
The third check runs on tgbn-genre, tgbn-trade, and tgbn-reddit, reporting NDCG@10 against Moving Average. Here we expect modest or no gains and say so.
The fourth check is a depth sweep from 2 to 64 layers on the PEMS datasets and on Chameleon and Squirrel, plotting accuracy, Dirichlet energy, and boundary radius together.
The fifth check measures cost: throughput and memory against THeGCN, because a curved model that is two orders of magnitude slower fails the deployment test the benchmark literature emphasizes.
The project succeeds if curvature beats the matched gate in the shock regime. It produces a clear negative result if it does not.
References
- H4G: Unlocking Faithful Inference for Zero-Shot Graph Learning in Hyperbolic Space
- Two Sides of the Same Coin: Heterophily and Oversmoothing in Graph Convolutional Neural Networks
- Dynamic Graph Prompting via Topology-Routed Mixed-Curvature Experts
- Fast Mixture of Curvature-Aware Experts for Diverse and Dynamic Graph Topologies
- Depth-Adaptive Graph Neural Networks via Learnable Bakry-'Emery Curvature
- Spectro-Riemannian Graph Neural Networks
- Neural Sheaf Diffusion: A Topological Perspective on Heterophily and Oversmoothing in GNNs
- THeGCN: Temporal Heterophilic Graph Convolutional Network
- Temporal Graph Benchmark for Machine Learning on Temporal Graphs
- Are we measuring oversmoothing in graph neural networks correctly?
- Hypformer: Exploring Efficient Transformer Fully in Hyperbolic Space
- Residual connections provably mitigate oversmoothing in graph neural networks
- PairNorm: Tackling Oversmoothing in GNNs
- Edge Sparsification via Temporal Forman-Ricci Curvature for Dynamic Graph Learning