acceptodds
Under review as a conference paper at ICLR 2027

Predictive Dynamic Graph World Modelling: From Temporal Context to Future Relations

Abstract

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 (a) *what has evolved*, a historical node-relation context encoder summarizes node-neighborhood trajectories and pair-specific histories. To characterize (b) *how it evolved*, a node-relation path-signature encoder captures ordered structural changes. To predict (c) *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.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Dynamic Graph, Joint-Embedding Predictive Architectures.

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_GcuXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DYG-WM, a dynamic graph learning framework inspired by the world modeling paradigm, aiming to predict future node and relation representations from historical graph observations. The framework consists of three main components. First, a historical node-relation context encoder captures node-neighborhood trajectories and pair-specific interaction histories. Second, a node-relation path-signature encoder models the ordered structural changes at both node and relation levels to characterize path-dependent graph evolution. Third, a dual-level joint-embedding predictive objective aligns predicted future node and relation representations with target representations obtained from an EMA-based encoder, enabling latent future-state prediction without explicitly reconstructing the target graph. The framework is evaluated on 12 dynamic link prediction datasets under different negative sampling strategies and two dynamic node classification datasets. Additional latent alignment analyses investigate the effectiveness of DYG-WM in learning predictive representations.

Critical strengths
  1. The paper explores a meaningful direction by extending predictive latent representation learning to dynamic graph evolution. Rather than focusing exclusively on downstream edge or node predictions, the proposed framework introduces an objective for jointly predicting future node and relation representations. This provides an interesting perspective on modeling graph evolution in latent space.
  2. The proposed architecture incorporates complementary historical information through node-neighborhood trajectories, pair-specific structural contexts, and path-signature representations of ordered structural changes. The introduction of relation modeling can potentially capture temporal dependencies that may not be adequately represented by node embeddings alone.
  3. The paper evaluates DYG-WM on both dynamic link prediction and dynamic node classification. Component ablations and latent alignment analyses also support the proposed architecture and its predictive learning objective.
Critical weaknesses
  1. Potential over-claim of the graph world modeling framing. Although the paper presents DYG-WM as a dynamic graph world model, the proposed framework more closely resembles temporal graph representation learning regularized by future latent prediction. The model inference is based on historical representations rather than explicit future graph state generation. The distinction between the proposed framework and self-supervised temporal graph learning remains insufficiently established, raising concerns about over-claiming.
  2. Limited empirical advantages and unclear contribution of the world modeling objective. In dynamic link prediction, DYG-WM does not demonstrate consistent or substantial improvements over strong baselines such as DyGFormer across different datasets and negative sampling settings. More importantly, the component ablations suggest that the predictive objective, which constitutes the central world modeling component, contributes less to the final performance than the path-signature encoder. Given that dynamic link prediction accounts for major part of empirical evaluation, the experimental evidence therefore does not sufficiently establish the distinctive benefits of the world modeling framing over conventional temporal graph learning approaches.
  3. Limited evidence of graph state evolution learning. The latent alignment experiments primarily demonstrate that the predicted representations become more similar to the target representations when the joint-embedding predictive objective is employed. However, improved latent similarity mainly verifies that the alignment objective is functioning as intended, rather than demonstrating that the model has learned meaningful or generalizable graph state evolution patterns.
Questions for authors
  1. Could authors clarify what distinguishes DYG-WM from conventional self-supervised temporal graph learning? It would be helpful to explain what specific world modeling capabilities the proposed framework provides beyond learning future temporal representations.
  2. Could authors provide further evidence of the practical benefits of the world modeling objective for dynamic link prediction, and provide ablation results on dynamic node classification?
  3. Could authors provide more evidence (the multi-step latent prediction for example) that the predicted latent representations capture meaningful graph evolution patterns beyond alignment with next-step embeddings?
6 comments

We thank the reviewer for finding the direction promising and the paper clearly written, and for questions that pinpoint exactly what a world-model claim must show. Our answers rest on new evidence:

  • The forecast captures how graphs evolve, not just alignment (W3/Q3). It identifies the true future pair state 4–8× above chance, up to three snapshots ahead. On UCI it beats a "no change" baseline by a margin that grows with the horizon: +5.2, +10.0 and +11.6 points.

  • The world-modelling objective has practical benefits for link prediction (W2/Q2). Its forecast alone scores links with AUC 84–92, adding it to the scorer improves Can. Parl. by +2.1 AUC, and JEPA-only pretraining beats a random encoder under a frozen probe.

  • Node-classification ablations are added (Q2).

  • The framing is scoped precisely: a latent, action-free world model of graph dynamics (W1/Q1).

W1/ Q1: DyG-WM is a world model in the JEPA sense: an action-free model that, from the history CtC_{t}, forecasts the future graph state in a learned latent space. Forecasting in latent space instead of generating the future graph is a deliberate JEPA design choice: it targets the predictable, abstract part of the future instead of reconstructing unpredictable details. We will state this scope explicitly, namely that there is no explicit graph generation and no actions or planning, and describe DyG-WM as a latent world model of graph dynamics.

Within this scope, DyG-WM differs from self-supervised temporal graph learning in what it learns and what it can do at inference time:

Method familyTraining signalExplicit forecast at inference?Pair-level future state?
Contrastive (e.g., CLDG)invariance across time spansnono
Masked reconstruction (e.g., MaskDGNN, DVGMAE)recover masked present structure / featuresnono
Node-level JEPA (SG-JEPA)next-snapshot node embeddingsnode level onlyno
DyG-WMhorizon-conditioned forecast of future node and relation latents vs. an EMA targetyes: y^ut\hat{y}_{u}^{t}, y^uvt\hat{y}_{uv}^{t} from CtC_{t} aloneyes

W2/ Q2: DyG-WM has the best average AUC-ROC rank in all three settings: 2.67 vs. 3.08 for DyGFormer (random negatives), 4.33 vs. 4.42 (historical) and 3.58 vs. 4.42 (inductive). Under AP the two are on par (Tables 8–9). Where DyGFormer is clearly ahead, the cause is dataset-specific:

  • Can. Parl.: events within a year are sorted by destination, and DyGFormer's un-masked padding then separates positive from negative batches.

  • MOOC: the dominant signal is a user's exact previous item, which no snapshot model observes. DyG-WM is the best snapshot method there.

Details are in our response to Reviewer qzG6 (W1). Revision: Contribution 3 is stated for AUC-ROC, and AP is discussed separately.

The signature and the objective play different roles, so their ablation drops are not directly comparable. The signature adds input information (the ordered structural changes of a pair), so removing it removes information. The objective shapes what the representation encodes, and its benefits for link prediction are direct:

  • A usable forecast. The forecast alone ranks links with AUC 84–92, while a model trained without the objective reaches the same supervised accuracy but has an uninformative forecast (AUC 37–58). Adding the forecast to the scorer improves Can. Parl. by +2.1 AUC / +1.9 AP on 4 of 5 seeds; on Enron the change is not significant (−0.6, p = 0.09). Replacing the observed relation context by the forecast costs only ~1 AUC.

  • No accuracy cost, small consistent gains. Under random negatives on US Legis., AP improves on all 5 seeds (+0.28). Elsewhere the differences are within seed variance.

  • Useful representations without labels. With JEPA-only pretraining and a frozen probe, DyG-WM beats an identically probed random encoder on Enron (+3.5 AUC, every seed; also above all self-supervised baselines of Table 1 in the same setting) and on UN Trade (+1.1 AUC / +2.3 AP, 4 of 5 seeds).

For the node classification ablation, on dblp, the three parts show the capability.

Macro-F1Micro-F1
full75.30±0.2873.30±0.31
w/o predictive objective73.63±0.5571.47±0.24
w/o path signature73.36±0.5071.32±0.37
w/o node trajectories73.40±0.6871.36±0.54
Reviewer_7MYsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes DyG-WM, an action-free predictive representation model for dynamic graphs. It combines historical snapshot GNN embeddings with recurrent summaries of node and neighborhood trajectories and second-order path signatures of structural descriptors. Exponential moving average (EMA) is leveraged to encode a future snapshot. Link prediction and node classification are done on top of the embeddings learned. Experiments are done to validate the effectiveness of the proposed method.

Critical strengths
  1. It is important to learn the features/relationships of nodes and edges in a dynamic graph.

  2. The experimental evaluation covers twelve link-prediction datasets, several negative-sampling strategies, a new-node setting, and two node-classification datasets.

  3. The ablation results indicate that designed modules contribute to downstream performance on the evaluated datasets.

Critical weaknesses
  1. The problem definition is not very clear. It is claimed to be a world model to learn the node features and relation of a dynamic graph. However, the evaluations are still conducted with comparisons in link prediction and node classifcation using learned node embeddings, making it difficult to distinguish with the literature of dynamic temporal graphs.

  2. In Lines 146-148, the dataset is partitioned "with roughly equal event counts, keeping equal-timestamp interactions in the same bin." How can we make sure these two requirements are satisfied simultaneously in a dynamic graph?

  3. In Lines 215-217, aggregation is done and mean pooling is further applied. In addition, the learned embeddings also go through an exponential moving average. Isn't going to be over-smoothed?

  4. Equation (10) seems to a combination of all the information without much design rationale.

  5. The performance of DyG-WM is comparable, or sometimes worse than DyGFormer, leading to concerns on the effectiveness of the design. In addition, it may be better to include SG-JEPA in link prediction as well.

Questions for authors
  1. Missing notation definitions. What are e_i and M in Line 145?

  2. In Lines 231-232, what is relation context and what is budget?

  3. In Line 249, what are d_i's? What does it mean by projected path increments?

  4. A line break may be needed for "Joint Objective" in Line 313?

4 comments

We thank the reviewer for the careful review and for recognising the importance of the problem, the breadth of the evaluation (12 link-prediction datasets, three negative-sampling protocols, the new-node setting and node classification) and the ablations showing that each module contributes. In short:

  • DyG-WM's forecast is explicit and testable (W1). It identifies the true future pair state 4–8× above chance, a capability that a model trained without the objective entirely lacks, at equal supervised accuracy.

  • The binning favours equal event counts on purpose (W2). Timestamp-aligned bins would leave coarse-timestamp datasets with too few snapshots for any snapshot model, and every causal guarantee needed by the query holds either way.

  • There is no over-smoothing (W3). The EMA acts on parameters, not embeddings, and target latents keep an effective rank of 59–77 out of 128.

  • Each term of Eq. (10) contributes information the others cannot (W4), as supported by ablations and controlled experiments.

  • DyG-WM has the best average AUC-ROC rank in all three settings, and DyGFormer's lead on two datasets is diagnosed (W5).

W1: The problem definition is not very clear. It is claimed to be a world model to learn the node features and relation of a dynamic graph. However, the evaluations are still conducted with comparisons in link prediction and node classifcation using learned node embeddings, making it difficult to distinguish with the literature of dynamic temporal graphs.

DyG-WM is a world model from the history CtC_{t} it forecasts the next graph state in latent space, at both the node level (y^ut{\widehat{y}}_{u}^{t}) and the relation level (y^uvt{\widehat{y}}_{uv}^{t}), and it is trained only by comparing these forecasts with the encoded future snapshot. This is what sets it apart from the methods in our comparison:

  • supervised link predictors learn a scoring function;

  • contrastive methods (e.g., CLDG) learn invariance across time;

  • reconstruction methods (e.g., MaskDGNN, DVGMAE) recover masked inputs.

None of them produces an explicit forecast of the future state.

This property can be evaluated directly, not only through downstream tasks. We test whether the forecast identifies the actual future: for each source uu, the observed destination and 7 random candidates are compared, and we check whether the forecast is closest to its own future target among the 8 targets.

TestCan. Parl.WikipediaUCI
Top-1 retrieval of the true future target (chance 12.5%), DyG-WM53.9%74.7%98.1%
Same, w/o predictive objective12.9%27.3%15.6%
AUC of the forecast–target distance alone as a link score, DyG-WM83.690.291.6
Same, w/o predictive objective36.757.638.6

W2: When the two requirements conflict, our protocol deliberately gives priority to equal event counts, for the following reasons. We will state this explicitly; the current sentence in Lines 146–148 is imprecise on this point.

  1. Timestamp-aligned bins would make the coarse-timestamp benchmarks unusable for any snapshot model. Can. Parl. has 14 distinct timestamps (one per year), UN Trade 32 and UN Vote 72. With timestamp-aligned bins, Can. Parl. would have 14 snapshots, i.e., 11 prediction windows for W=4W = 4, and only about two test windows after the 70/15/15 split. Equal-count bins give every dataset 50 snapshots and the same chronological split.

  2. Equal counts keep snapshots comparable. Every context step summarises the same number of interactions, so the signature increments have the same scale over time and across datasets. Timestamp-aligned bins would vary by up to 14× within one dataset (UN Vote: 1,775 to 24,432 events per timestamp; Can. Parl.: 3,098 to 8,334).

  3. All snapshot methods use identical bins. CLDG, MaskDGNN, DVGMAE and DyG-WM consume the same snapshots, windows and splits. All methods, event-level and snapshot-level, are scored on exactly the same test events and negatives.

  4. The causal guarantees that matter for the query hold regardless of binning.

    • The target bin is never in the context: CtC_{t} contains bins t−W+1,…,t−1t - W + 1,\ldots,t - 1 only.

    • The event-level pair history used by the scorer contains only interactions strictly before the query time τ\tau.

    • Earlier snapshots can share a timestamp with a query only when the dataset's own timestamp resolution is coarser than one bin.

  5. On fine-grained datasets the two requirements coincide. No test query shares a timestamp with its context on UCI, LastFM and MOOC, and fewer than 5% do on Contacts and Enron.

W3: No. The EMA does not smooth representations: it acts on the target encoder's parameters, not on embeddings. The target encoder's weights are an exponential moving average of the online weights, θˉ←ηθˉ+(1−η)θ\bar{\theta} \leftarrow \eta\bar{\theta} + (1 - \eta)\theta, and it only provides regression targets. No embedding is ever averaged over time or over training steps.

Mean pooling is restricted to one-hop neighbourhood trajectories and never replaces a node's own state. Eq. (2) forms a separate neighbourhood trajectory (hˉus)s(\bar{h}_{u}^{s})_{s}. The node's own trajectory (hus)s(h_{u}^{s})_{s} has its own recurrent encoder, and both are combined with the latest node state through gated residual fusion (Eq. 3), so the gate can down-weight an uninformative neighbourhood. The GNN has 2 layers.

The learned representations confirm this:

  • the target latents have an effective rank of 58.7, 77.3 and 71.7 out of 128 dimensions on Can. Parl., Wikipedia and UCI;
  • they separate a node's true future partner from random candidates (W1 table), which near-identical embeddings could not do;
  • removing the historical context lowers AUC (Table 4: −3.5 / −2.1 / −2.3 on Enron / Can. Parl. / LastFM).

W4: Eq. (10) combines three sources that each answer a different question about a pair (u,v)(u,v); none can be derived from the others.

TermWhat it encodesWhy the other terms cannot provide it
gθ(χ(cut,cvt))g_{\theta}(\chi(c_{u}^{t},c_{v}^{t})), χ(a,b)=[a∥b∥a⊙b∥a−b]\chi(a,b) = \lbrack a \parallel b \parallel a \odot b \parallel a - b\rbrackcurrent states and histories of the two endpoints and their compatibility; χ\chi is ordered, so direction is preserved—
guvtg_{uv}^{t} (relation-centric context, Eq. 4)history of the pair's shared surroundings: bridging nodes and recent common neighboursendpoint states are computed independently for uu and vv and do not record what they share
SuvtS_{uv}^{t} (pair path signature, Eq. 5)ordered trajectory of the pair's structural statisticsrecurrent summaries do not explicitly preserve the order and interaction of changes

W5: DyG-WM has the best average AUC-ROC rank in every link-prediction setting: 2.67 vs. 3.08 for DyGFormer (random negatives), 4.33 vs. 4.42 (historical) and 3.58 vs. 4.42 (inductive). Under AP the two are on par (Tables 8–9). The datasets where DyGFormer is clearly ahead have dataset-specific causes:

  • Can. Parl.: events within a year are sorted by destination, and DyGFormer's un-masked padding then differs between positive and negative batches. That statistic alone separates them with AUC 100, and it drops to chance once the order within a timestamp is shuffled.
  • MOOC: the dominant signal is a user's exact previous item (AUC 0.95). Every snapshot model sees it only at bin granularity (0.71). Among snapshot methods, DyG-WM is the best by a wide margin.

SG-JEPA was proposed and evaluated only for node classification, which is why we compare with it there (Table 3). Its paper reports DBLP, Tmall and Patent node classification, defines no link-prediction protocol or pair scorer, and its objective forecasts node-level embeddings only. A link-prediction SG-JEPA would therefore be our own design (readout, pair scorer, protocol) and would reflect those choices rather than the authors' method. The comparison the reviewer has in mind is still covered:

  • Self-supervised snapshot encoders evaluated through a frozen encoder and a pair probe are represented in Table 1 by CLDG, MaskDGNN and DVGMAE. DyG-WM ranks far above all three in every setting: average AUC rank 2.67 / 4.33 / 3.58 vs. at best 7.75 / 6.50 / 7.42.
  • DyG-WM forecasts the future pair state, which a node-level JEPA objective does not train, and this forecast is 4–8× above chance (W1).

Q1–Q4. Notation. Thank you; all four will be fixed in the revision.

  • Q1. eie_{i} is the feature vector of interaction ii (the DyGLib edge features, e.g., 172-dimensional text features on Wikipedia, the scalar vote weight on Can. Parl.). MM is the total number of interactions. Both are defined at first use.

  • Q2. The relation context of (u,v)(u,v) at bin tt is the node set KuvtK_{uv}^{t} pooled in Eq. (4). It always contains uu and vv and is filled with their historical neighbours, ranked by recency plus a bridging bonus for nodes adjacent to both. The budget BB is its maximum size (B=32B = 32 throughout), which bounds the per-pair cost and keeps the most relevant shared neighbours.

  • Q3. Each context snapshot provides a vector of pair statistics: direct link, endpoint degrees, common neighbours, Jaccard, activity, coverage, and a time channel. The path has L=W−1L = W - 1 steps. Step 1 is the statistics at the first context snapshot, and step i≥2i \geq 2 is the change since the previous one. "Projected" means each step is mapped by a learned linear layer with tanh into RdsR^{d_{s}} (ds=16d_{s} = 16); these are d1,…,dLd_{1},\ldots,d_{L} in Eq. (5). The statistics are listed in the appendix.

  • Q4. "Joint Objective" will start a new paragraph.

Reviewer_qzG6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper frames dynamic graph learning as a world-model problem: rather than treating node labels or observed edges as the primary supervision, it uses historical graph states to predict future node and relation states directly, without reconstructing the full future graph. The model first organizes historical node and relation context, then encodes the ordered sequence of structural changes, and learns representations that predict both future node and relation states. Comparative results across several link-prediction and node-classification datasets indicate that this direction is useful for both tasks.

Critical strengths
  1. Broad system-level evaluation across 12 link-prediction datasets and two node-classification datasets. Tables 1 and 2 report DYG-WM average AUC-ROC ranks of 2.67, 4.33, and 3.58 under the three transductive negative-sampling regimes, plus 2.83 in the new-node setting, all best among the compared methods; Table 3 shows the highest reported F1 means across the three labeled-node ratios on DBLP and Tmall.
  2. The main training and inference procedures are documented. Appendices A.5 through A.8 give pseudocode for representation construction, dual-level predictive training, and dynamic link inference, and Appendix B states time and space complexity.
Critical weaknesses
  1. The headline contribution claim is not matched by the paper's own appendix tables. Contribution 3 states that DYG-WM achieves the best average rank in every evaluated link-prediction setting, but the AP average ranks in Table 8 are better for DyGFormer under both random and historical negative sampling, and Table 9 ties DyGFormer and DYG-WM in the new-node AP setting. The large per-dataset gaps on datasets such as Canadian Parliament and MOOC are also not diagnosed in the text.
  2. The independent contribution of the core mechanism has not been separated by controlled experiments. The relation target is obtained by an ordered transformation of endpoint target latents rather than being an independently observed pair-level state; deployed link scoring uses the online encoder output rather than the predictor output; and removing the predictive objective gives comparatively small drops on the three ablated datasets, where mean plus or minus one-standard-deviation intervals overlap pairwise, so formal statistical testing is needed before claiming non-significance. Removing the entire path-signature module produces a larger drop, but without a control that keeps the same structural descriptors, time channels, supervision, and parameter budget while replacing the second-order term with a first-order or order-shuffled version, it is difficult to separate order information from descriptor richness and capacity.
  3. Whether the node-classification probe budget is matched across methods needs clarification. Appendix A.2 specifies that DYG-WM fits separate MLP probes to the predicted and online representations and averages their logits, but it does not state per baseline how many probes or what logit fusion the unified pipeline uses. Whether the uniformly best F1 numbers in Table 3 are obtained under a matched probe or ensemble budget is an attribution question the authors should address directly.
  4. The latent-alignment heatmaps provide limited evidence. In Figure 3, several target rows show a block-like pattern of high cosine similarity, and Figures 6 and 7 show a related pattern; high cosine values alone do not establish node-specific prediction and cannot rule out alignment along a shared direction, yet the paper reads the elevated cosines as evidence that the predictive objective is effective.
Questions for authors
  1. Could Contribution 3 be restricted to the AUC-ROC setting with the AP results discussed separately? What diagnostic explains the larger gaps on Canadian Parliament, MOOC, and related datasets?
  2. With the same structural descriptors, time channels, supervision, and parameter budget, how do a first-order-only signature, a time-shuffled path, and the full second-order signature compare?
  3. How do results compare between a node-only predictive objective and the full dual-level objective, and what representation quality results from JEPA-only pretraining followed by a probe with link and ranking supervision removed?
  4. Current inference uses the online representation; how would results change if scoring used the predictor output instead?
  5. Could you clarify the number of probes and logit-fusion rule used for each baseline under the node-classification protocol, and report DYG-WM versus baselines under matched probe or ensemble budgets?
4 comments

We thank the reviewer for the detailed and constructive review. We ran every requested control on the same hardware and protocol as the paper (historical negatives at test time, paired seeds). In short:

  • Order in the signature matters, and the ordered term is the only part of level 2 that helps (Q2). Same descriptors, time channels and parameter count; Enron p ≤ 0.022.
  • The predictor produces a real, pair-specific forecast (W4, Q4). It retrieves the true future pair state 4–8× above chance (54–98% vs. 12.5%). Scoring from the forecast alone gives AUC 84–92, and replacing the observed context by the forecast costs only ~1 AUC.
  • Contribution 3 will be stated for AUC-ROC, where DyG-WM has the best average rank in all three settings; AP is discussed separately (W1).
  • The Can. Parl. and MOOC gaps are diagnosed: a batch-padding artefact of DyGFormer on Can. Parl., and the snapshot granularity on MOOC (W1).
  • The node-classification probe protocol is fully specified and a matched single-probe comparison is added (W3/Q5).

W1: Yes, Contribution 3 will be stated for AUC-ROC, which is what our results support. DyG-WM has the best average AUC-ROC rank in every link-prediction setting: 2.67 vs. 3.08 for DyGFormer with random negatives, 4.33 vs. 4.42 with historical negatives and 3.58 vs. 4.42 in the inductive setting. Under AP the two methods are on par (2.67 vs. 2.42 and 3.83 vs. 3.67, Table 8; tied at 2.58 in the new-node setting, Table 9).

Revision: "DyG-WM achieves the best average AUC-ROC rank in every evaluated link-prediction setting and is on par with the strongest baseline under AP." AP gets its own paragraph in Sec. 4.1.

Can. Parl. DyGFormer’s high performance appears to come from batch construction rather than temporal modeling. Events have yearly timestamps and are sorted by destination within each year. DyGLib processes consecutive batches of 200 events, so positive samples tend to share a destination hub, while negatives use random destinations. DyGFormer pads destination sequences to the batch maximum and mean-pools without masking. Consequently, the number of destination patches alone achieves 100% AUC and AP on validation and test, but AUC drops to 54% after shuffling events within timestamps, which preserves causal information. Other baselines use fixed-size neighbor sampling and are unaffected, consistent with DyGFormer being the only method above 90 (97 vs. 65–84), also in the original DyGLib benchmark. The causal signal is weaker: 40–57% of positive pairs are new each year, roughly half of destination hubs change, and a gradient-boosted model using 25 causal features reaches only 71–78% AP. Validation also spans the 2015 federal election.

MOOC. The strongest predictive signal is unavailable to snapshot models: a user's immediately previous item yields AUC 0.95 at the event level, but only 0.71 using the previous snapshot's last item, since 87% of events have an earlier same-user event within the current snapshot. Event-level models retain this information. Among snapshot methods, DyG-WM performs best by a wide margin (AP 81.94 vs. 72.35 for MaskDGNN, 68.41 for DVGMAE, and 64.90 for CLDG).

W2: (a) Relation target. The target is pair-specific even though it is built from endpoint latents. The target is yuv=gψˉ([zu,zv,zu⊙zv,zu−zv])y_{uv} = g_{\bar{\psi}}(\lbrack z_{u},z_{v},z_{u} \odot z_{v},z_{u} - z_{v}\rbrack) where zu,zvz_{u},z_{v} are the EMA encoder's latents of the endpoints on the future snapshot. Because they are computed by message passing over that snapshot, the target reflects the future structure around both endpoints, and its directed construction distinguishes (u,v)(u,v) from (u,v′)(u,v'). The forecast indeed recovers the specific pair: it retrieves its own future target among 8 candidates in 54–98% of cases (chance 12.5%, W4).

(b) Scoring from the predictor output. Done; see Q4. The forecast alone carries most of the scoring information.

(c) Statistical support for the predictive objective. The paired test shows that the objective leaves supervised accuracy intact (never significantly lower) and is what produces the forecasting capability:

Accuracy. Under random negatives on US Legis., AP improves on all 5 seeds (+0.28; AUC +0.14, p = 0.059). On Enron and Can. Parl. with historical negatives the differences are within seed variance (paired p = 0.74 / 0.73).

Forecasting. With the objective, the prediction retrieves its own future pair state 4–8× above chance and scores links on its own with AUC 84–92. Without it, both fall to chance (W4). Since the two models reach the same supervised accuracy, this difference is due to the objective, not to capacity or supervision.

Label-free representations. In node classification the encoder is trained by the objective alone (W3/Q5). With JEPA-only pretraining, a frozen link probe beats an identically probed random encoder on Enron (every seed) and UN Trade (4 of 5 seeds) (Q3).

(d) Order vs. descriptor richness and capacity. Done; see Q2. With identical descriptors, time channels, supervision and parameter count, removing the order-dependent term lowers AUC on Enron on every seed (p ≤ 0.022).

W3: For the matched comparison, DyG-WM follows the baseline rule exactly: one probe, on the better of its two learned views (predicted or encoder latent) by validation Macro-F1, and no 2-hop view. The encoder view is selected on every seed.

DBLP, 40% labelsProbesMacro-F1Micro-F1
SG-JEPA (Table 3)173.96 ± 0.5974.20 ± 0.55
GraphMixer (Table 3)171.69 ± 0.7972.15 ± 0.67
MaskDGNN (Table 3)170.98 ± 0.8471.01 ± 0.67
TCL (Table 3)169.76 ± 1.0970.35 ± 1.55
DyG-WM172.70 ± 0.8272.83 ± 0.44
DyG-WM375.45 ± 0.3575.79 ± 0.24

It is above all ten non-JEPA baselines, by +1.0 Macro-F1 over GraphMixer and +1.7 over MaskDGNN, and 1.3 points below SG-JEPA. The multi-view readout adds about 2.5–3 points. It combines the predicted latent, the encoder latent and the 2-hop feature view, and it is what places DyG-WM first in Table 3.

Q1: See W1.

Q2: First-order only: level 2 set to zero, so only the net change S(1)=diS^{(1)} = d_{i} remains.

Order-free: level 2 replaced by its average over all orderings of the increments, 12S(1)⊗S(1)\tfrac12 S^{(1)}\otimes S^{(1)}. This is exactly what a uniformly time-shuffled path gives in expectation, without sampling noise. The Lévy area, the only order-dependent quantity at depth 2, becomes zero, and every order-independent second-order feature is kept.

Signature (AUC, historical negatives)EnronCan. Parl.LastFM
First-order only72.84 ± 2.2676.59 ± 4.2175.11 ± 0.45
Time Shuffle73.12 ± 2.8975.63 ± 4.2775.33 ± 0.58
Full second-order (DyG-WM)74.94 ± 1.9777.01 ± 2.4277.60 ± 0.53

Q3: First-order only: level 2 set to zero, so only the net change S(1)=diS^{(1)} = d_{i} remains.

Order-free: level 2 replaced by its average over all orderings of the increments, 12S(1)⊗S(1)\tfrac12 S^{(1)}\otimes S^{(1)}. This is exactly what a uniformly time-shuffled path gives in expectation, without sampling noise. The Lévy area, the only order-dependent quantity at depth 2, becomes zero, and every order-independent second-order feature is kept.

Signature (AUC, historical negatives)EnronCan. Parl.LastFM
First-order only72.84 ± 2.2676.59 ± 4.2175.11 ± 0.45
Time Shuffle73.12 ± 2.8975.63 ± 4.2775.33 ± 0.58
Full second-order (DyG-WM)74.94 ± 1.9777.01 ± 2.4277.60 ± 0.53

Q3: How do results compare between a node-only predictive objective and the full dual-level objective, and what representation quality results from JEPA-only pretraining followed by a probe with link and ranking supervision removed?

By design, the predictive levels are matched to the task, and both configurations the reviewer asks about are already the ones we use (Appendix A, Table 7):

  • Node classification uses the node-level objective only. The relation and link objectives are disabled (λN=1.0\lambda_{N} = 1.0, λR=0\lambda_{R} = 0). The encoder is trained without labels and frozen, and only then is a probe fitted. This is exactly node-only, JEPA-only pretraining followed by a probe, and it yields the best node-classification results in Table 3.

  • Link prediction uses both levels together with the link losses (λN=λR=0.02\lambda_{N} = \lambda_{R} = 0.02, λlink=λrank=1\lambda_{link} = \lambda_{rank} = 1). Here the target is a pair, so the relation level is needed: it trains the relation predictor whose pair-level forecast the link task relies on.

Q4: The default scorer reads [ruv,ouv,quv]\lbrack r_{uv},o_{uv},q_{uv}\rbrack, with quvq_{uv} the observed relational context.

Replace uses [ruv,ouv,y^uv]\lbrack r_{uv},o_{uv},{\widehat{y}}_{uv}\rbrack, so the relational context reaches the score only through the predictor; the parameter count is unchanged.

Append uses [ruv,ouv,quv,y^uv]\lbrack r_{uv},o_{uv},q_{uv},{\widehat{y}}_{uv}\rbrack.

Scorer inputEnron AUCEnron AP
Replace: [r,o,y^]\lbrack r,o,\widehat{y}\rbrack73.81 ± 1.8175.62
Append: [r,o,q,y^]\lbrack r,o,q,\widehat{y}\rbrack74.33 ± 1.8676.40
DyG-WM: [r,o,q]\lbrack r,o,q\rbrack74.94 ± 1.9776.78

Q5: See W3.

Comments on the paper

2 comments

Dear reviewers,

The rebuttal period is ending soon. Please read the authors rebuttal and check if they addressed your concerns. Please acknowledge the rebuttal and adjust your scores accordingly.

Thanks.

-AC

AuthorsOpenReview ↗

Common Response to All Reviewers

We thank all reviewers for their careful reading and helpful comments. We are glad that the reviewers found several aspects of the work valuable:

  1. The direction itself: extending predictive latent representation learning to how dynamic graphs evolve, by forecasting both future node and future relation representations (Reviewers GcuX, 7MYs, qzG6).
  2. The architecture, which combines node-neighbourhood trajectories, pair-specific structural context and path signatures of ordered structural changes. Modelling relations explicitly can capture temporal dependencies that node embeddings alone may miss (Reviewer GcuX).
  3. The breadth of the evaluation: 12 link-prediction datasets, several negative-sampling strategies, the new-node setting and two node-classification datasets, where DyG-WM has the best average AUC-ROC ranks and the highest F1 means (Reviewers qzG6, 7MYs, GcuX).
  4. The ablations and latent alignment analyses, which support both the modules and the predictive objective (Reviewers 7MYs, GcuX).
  5. The documentation of training and inference, including pseudocode and a complexity analysis (Reviewer qzG6).

Reviewer GcuX described the work as 'a meaningful direction' that offers 'an interesting perspective on modeling graph evolution in latent space', and Reviewer qzG6 noted that the results 'indicate that this direction is useful for both tasks'. Reviewer 7MYs also observed that 'the designed modules contribute to downstream performance on the evaluated datasets'. We appreciate these comments.

Our responses to each reviewer follow below. We have run the additional experiments that were requested and will add them, together with the clarifications, to the revision. Please let us know if anything remains unclear.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.