Do Graph Event Models Give Every Node the Same Clock?

Many event datasets live on a graph. Users message each other, forum users reply within linked communities, collisions occur across adjacent boroughs. Models for such data must predict which node acts next and how long until it does. A common recipe pairs two parts. A graph neural network summarizes each node's relational history by repeatedly averaging over neighbors. An "intensity-free" temporal point process head then turns that summary into a probability density over the waiting time to the next event, typically a mixture of log-normal distributions whose likelihood, quantiles and samples are all available in closed form. Graph learning has a well-known failure mode called over-smoothing: as averaging layers are stacked, node representations drift toward one another until the nodes can no longer be told apart. This note asks a simple question that nobody in the literature we found has measured. When the encoder over-smooths, do the predicted waiting-time distributions also collapse, so that every node is assigned roughly the same clock?

We propose to measure over-smoothing where it actually costs something, in the space of predicted timing densities, rather than in embedding space. Log-normal mixtures make this unusually cheap. In log-time each prediction is a Gaussian mixture, so distances between two nodes' predicted densities, per-node calibration, and a "swap test" that scores one node's true next interval under another node's density are all exact closed-form quantities. With these tools we will test whether the standard embedding diagnostic (Dirichlet energy) predicts timing collapse. We will also test whether collapse explains two known weaknesses of intensity-free models: poor fit of type-wise distributions when there are many event types, and poor arrival-time RMSE. Finally, we will check whether a fix aimed directly at timing collapse beats generic anti-over-smoothing machinery.

Core Claim

The claim has three parts. First, in graph-coupled intensity-free event models, over-smoothing shows up as a loss of node-specific timing information. Pooled calibration and pooled likelihood can hide this loss. Second, embedding-space measures cannot separate harmless agreement (neighboring nodes that genuinely share timing behavior) from harmful collapse (nodes with different true clocks that receive the same prediction). Are we measuring oversmoothing in graph neural networks correctly? raises exactly this concern for static GNNs, and density-space diagnostics resolve it because they can be checked against observed intervals. Third, a closed-form contrastive penalty on the swap test is a targeted remedy. It should recover node-specific timing at least as well as replacing the encoder with a non-dissipative backbone, and more cheaply.

What would be new is that over-smoothing is defined, measured and corrected through the output densities of a generative event model. No paper in our search does this. Dynamic Latent Graph-Guided Neural Temporal Point Processes (VAETPP) couples a GNN and graph-recurrent decoder to a log-normal mixture head but never measures representation collapse. The two closest efforts to fight collapse in dynamic graphs are Long Range Propagation on Continuous-Time Dynamic Graphs (CTAN) and On Vanishing Gradients, Over-Smoothing, and Over-Squashing in GNNs (GNN-SSM). Both are discriminative and contain no event density at all.

The Measurements

Write $p_u(\cdot \mid \mathcal{H}_t)$ for node $u$'s predicted density over $x = \log \tau$, where $\tau$ is the waiting time to $u$'s next event. Under a log-normal mixture head this is a Gaussian mixture with weights $w_i$, means $\mu_i$ and variances $\sigma_i^2$. For a second node $v$ with components $(v_j, \nu_j, s_j^2)$, the overlap

$$\int p_u(x)\,p_v(x)\,dx = \sum_{i}\sum_{j} w_i v_j\, \mathcal{N}(\mu_i;\ \nu_j,\ \sigma_i^2 + s_j^2)$$

is exact. The squared $L_2$ distance $D(u,v)$ between the two predicted clocks therefore costs a few matrix operations.

Ground truth is known on synthetic data, so there we report a collapse ratio: predicted $D$ over true $D$, averaged across node pairs. On real data, ground truth is unavailable, so we use the swap gap

$$G = \mathbb{E}_{(u,\tau),\,v}\big[\log p_u(\log\tau) - \log p_v(\log\tau)\big],$$

where $(u,\tau)$ is a held-out event, $v$ is a node drawn at random from the same snapshot, and the expectation runs over held-out events. A collapsed model has $G \approx 0$ because swapping clocks costs nothing. A model that uses node-specific timing information has large positive $G$. Harmless agreement also gives $G \approx 0$, but in that case it loses no likelihood. The two cases are separated by comparing $G$ against the gain in held-out likelihood that a per-node model achieves on the same data.

We add a second real-data check based on per-node calibration. The probability integral transform $F_u(\tau)$ uses the closed-form mixture CDF. Collapse predicts that the pooled transform looks uniform while its spread across node groups is large: some nodes are systematically early and others systematically late.

Research Program

The study has four pieces.

The first is a controlled synthetic benchmark modeled on CTAN's T-PathGraph. Each node's true waiting-time law is a log-normal mixture whose mode locations are set by the features of a source node $k$ hops away. Sibling nodes in the same community can therefore have very different clocks. Varying $k$ from 1 to 20 sets how much depth an encoder needs, and adding a "consensus" variant, where neighbors truly share clocks, tests whether the diagnostics wrongly flag harmless agreement.

The second is a backbone sweep. We hold one log-normal mixture head, taken from Intensity-Free Learning of Temporal Point Processes, fixed across five history encoders:

  • stacked GCN-style message passing at increasing depth;
  • the same stack with PairNorm (PairNorm: Tackling Oversmoothing in GNNs);
  • a filterbank encoder with high-pass diversification channels (Complete the Missing Half);
  • a CTAN-style anti-symmetric continuous-time encoder;
  • a GNN-SSM-style encoder with transitions near the edge of stability.

For every model and depth we log Dirichlet energy, collapse ratio or swap gap, per-node calibration spread, type-wise NLL and RMSE.

The third is the targeted fix. We add a hinge penalty to the training loss that requires the swap gap on training events to exceed a margin. It is computed with the same closed-form densities, adds negligible cost, and is applied to the cheap GCN-style encoder.

The fourth is an honest decomposition. We regress RMSE and type-wise NLL on the collapse measures and on Dirichlet energy. The point is to test whether collapse explains part of the arrival-time deficit that EasyTPP and An Extensive Survey With Empirical Studies on Deep Temporal Point Process report. The second of these attributes goodness-of-fit to the density head and type accuracy to the history encoder. Our hypothesis is that type-wise timing fit also depends on the encoder, through collapse.

Baselines

The naive reference point is a shared log-normal mixture head driven by an encoder of zero or one hop, which ignores relational history beyond the node itself. Any graph encoder must beat it on held-out NLL to justify its existence. We also report a homogeneous Poisson rate per node, the $n/t_n$ estimator used in CorDGT, as the parameter-free floor.

The strongest existing graph-coupled intensity-free model is VAETPP. It already combines a GNN encoder, a latent dependency graph and a log-normal mixture timing head. On NYMVC it reports test NLL of −5.952, against −2.016 for static-graph models and −1.889 for the Transformer Hawkes Process. On MathOF it reports next-type accuracy of 0.321, against 0.286 for the best baseline. We will rerun the released configuration on these two datasets, report our reproduced numbers alongside the published ones, and then apply our swap-gap penalty to the same model. Beating it means lower NLL than our reproduction on NYMVC, with type accuracy no worse on MathOF, under an identical hyperparameter search budget.

The strongest non-graph competitor is FlexTPP from Transformers for Mixed-type Event Sequences, an autoregressive Transformer with a spline-flow time head. The file says it set the NLL record on 4 of 5 EasyTPP datasets but gives no per-dataset values, so we have no reported number on Lacuna for FlexTPP NLL. We will run it ourselves on Retweet, StackOverflow and Taobao.

The strongest anti-collapse encoder is CTAN, which has no density head. For a fair comparison we attach the same log-normal mixture head. On Wikipedia and Reddit, the file contains no reported TPP NLL for CTAN, and no reported RMSE for any model.

Every model gets the same splits, the same head with the same number of mixture components, the same compute cap per run, and the same number of tuning trials. The key ablation isolates the penalty: a GCN-style encoder at depth 8 with the penalty, compared against the same encoder without it, with PairNorm, and against the CTAN-style encoder without the penalty.

Why not just…?

Why not just give each event type or node its own head offset? A learned per-node bias on the mixture parameters would trivially restore node-specific clocks on transductive data, and it is the most serious threat to the targeted fix. It cannot generalize to unseen nodes, and it cannot express timing differences that come from a node's current neighborhood rather than its identity. Our synthetic benchmark is built so that siblings share identity-level statistics but differ in their $k$-hop sources. We will also evaluate inductively on Wikipedia and Reddit with held-out nodes. If the offset matches the penalty in both settings, we will conclude that collapse is a head problem, not an encoder problem, and say so.

Why not just keep the encoder shallow? hpGAT shows shallow attention with injected high-order proximity widens the receptive field without deep averaging. CTAN reports that shallow and memory-based baselines (TGN, JODIE) collapse on T-PathGraph as path length grows toward 20. Our synthetic sweep over $k$ tests the same question for timing densities. If shallow models match deep ones on real data, the honest reading is that real benchmarks do not need depth, and that finding is itself worth reporting.

Why not just add PairNorm or swap in a non-dissipative backbone? These approaches keep embedding variance high. Embedding variance is not the same thing as distinct, correct clocks, and GNN-SSM argues that normalization does not remove contractive dynamics. The backbone sweep reports Dirichlet energy and swap gap side by side. If high energy coexists with near-zero swap gap, generic fixes are aimed at the wrong target. If CTAN-style encoders already reach a high swap gap and the penalty adds nothing, the targeted fix is redundant, and we will report that.

Why not just use a Transformer over flattened events? FlexTPP is strong on NLL. Attention has its own token over-smoothing, as studied in Setting the Record Straight on Transformer Oversmoothing and Demystifying Oversmoothing in Attention-Based Graph Neural Networks, and the swap gap applies to it unchanged. We will compute the swap gap for FlexTPP, which makes this a measured comparison rather than an argument.

Proposed Experiments and Decisive Checks

The study runs on four groups of data: the synthetic $k$-hop benchmark; NYMVC and MathOF, as in VAETPP; Wikipedia and Reddit, used by both CTAN and the intensity-free TPP paper; and Retweet, StackOverflow and Taobao from the EasyTPP suite, using a VAETPP-style learned type graph. Metrics are held-out NLL, type-wise NLL, RMSE of the mixture mean, swap gap, per-node calibration spread and Dirichlet energy.

Three results would decide the proposal.

First, on synthetic data, the collapse ratio should fall with GCN depth well before held-out pooled NLL degrades. It should stay near 1 on the consensus variant, which shows the diagnostic does not flag harmless agreement. If it does not fall with depth, the phenomenon does not exist and the project stops there.

Second, across all models and datasets, the swap gap should predict type-wise NLL and RMSE better than Dirichlet energy does. If energy predicts as well, the density-space view adds nothing over existing tools.

Third, the penalized shallow-cost encoder should match the CTAN-style encoder's swap gap and NLL at a fraction of its cost, and should improve on our VAETPP reproduction on NYMVC. If CTAN plus the head dominates, the contribution narrows to the diagnostic and the evidence that non-dissipative backbones help generative event models. That is a smaller result, but still a real one.