Dynamic Graph World Modelling: Learning Future Dynamics from Temporal Context
DyG-WM tackles dynamic graph world modelling by predicting future latent node and relation states from temporal context, rather than fitting specific supervised tasks or reconstructing raw edges. The core idea is a dual-level joint-embedding predictive architecture that combines historical node-relation context encoders with path-signature encoders to capture ordered structural changes. This matters because it aims to generalize across both link prediction and node classification without expensive task-specific labels.
The paper presents a well-structured and empirically thorough contribution that advances self-supervised dynamic graph learning through a carefully designed predictive latent-space objective. The three-component design is conceptually coherent and the 12-dataset link-prediction evaluation provides broad coverage. However, the 'world model' framing is more aspirational than demonstrated: the method predicts future snapshot embeddings but does not show planning, decision-making, or multi-step rollouts, and the authors themselves concede it is 'an action-free relational transition model rather than a full planning system.' The strongest results are on link prediction, while the node-level improvements over the closely related SG-JEPA are claimed but evaluated on only two datasets, leaving open questions about scalability to the very large graphs that SG-JEPA targets.
The technical design holds up well. The historical node-relation context encoder explicitly models pair-specific structural history beyond endpoint representations, addressing a real gap in node-centric DGNNs. The use of second-order path signatures to encode ordered structural changes is a principled choice, and the ablation study confirms that removing path signatures causes the largest performance drops. The dual-level JEPA objective avoids reconstruction and contrastive augmentation, which is a sensible direction for sparse dynamic graphs. The unified DyGLib-based reproduction pipeline and reported standard deviations over five seeds strengthen empirical credibility.
The 'world model' claim is the most significant conceptual overreach. Predicting latent embeddings one step ahead is not equivalent to learning a transition model suitable for planning or counterfactual reasoning; without action conditioning or multi-step rollout experiments, calling this a 'world model' is misleading. The relation-level predictive target is constructed from the target snapshot using an EMA encoder and endpoint latents, but the paper does not verify whether these relation latents actually capture meaningful relational dynamics beyond what endpoint similarity would yield. Comparisons against SG-JEPA are limited to node classification on two datasets, even though SG-JEPA is described as the most closely related method and scales to 13 million edges—DyG-WM reports no experiments on similarly large graphs or training-efficiency comparisons with SG-JEPA. Finally, several design choices, such as the relation-context budget $B=32$ and the extensive loss-weight tuning, are dataset-dependent and validated only on small hold-out sets, raising questions about transferability.
The evidence largely supports the link-prediction superiority claims: DyG-WM achieves the best average rank across three transductive negative-sampling settings and the new-node inductive setting, with concrete AUC-ROC margins on many datasets. However, the comparison with SG-JEPA is too narrow to support the claim of being the first or best joint-embedding predictive architecture for dynamic graphs generally; SG-JEPA is only evaluated on node classification, where DyG-WM wins on two medium-scale datasets, but SG-JEPA's primary contribution is scalability to 13-million-edge dynamic graphs, a regime DyG-WM does not probe. The efficiency analysis in Figure 4 omits SG-JEPA entirely. The robustness analysis is encouraging but limited to two datasets and feature noise only, not structural noise or distribution shift.
Reproducibility is reasonably strong. The authors provide a GitHub repository, implement all baselines in a unified DyGLib pipeline, and report detailed hyperparameters in Appendix A.2, including EMA coefficients, loss weights, and learning rates. The 5-seed standard deviations are reported for all main results. Key remaining gaps are the lack of dataset-specific ablations for the window size $W$ and relation-context budget $B$, the absence of wall-clock time or memory measurements against SG-JEPA, and the fact that the node-classification pipeline freezes encoders and averages two MLP probes, which adds a procedural step not fully justified by ablation.
Dynamic graph learning models how entities and their relations evolve from historical observations to predict future graph states. Existing methods do not fully address dynamic graph world modelling, which requires learning transitions from temporal context to future latent graph states. This is difficult for three reasons: nodes and relations rely on different historical evidence, structural evolution is path-dependent, and future snapshots provide sparse and realization-specific targets. We propose Dynamic Graph World Modelling (DyG-WM), the first framework to jointly predict future node and relation representations from temporal context by answering three questions: what has evolved, how it evolved, and what will evolve next. To capture what has evolved, a historical node-relation context encoder summarizes node-neighborhood trajectories and pair-specific histories. To characterize how it evolved, a node-relation path-signature encoder captures ordered structural changes. To predict what will evolve next, a dual-level joint-embedding objective predicts target-snapshot node and relation representations without reconstructing the raw graph. Experiments on 12 link-prediction and two node-classification datasets demonstrate the effectiveness of DyG-WM across both tasks.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.