Pith. sign in

REVIEW 4 major objections 5 minor 34 references

T-GRAB: A Synthetic Diagnostic Benchmark for Learning on Temporal Graphs

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read T-GRAB claims current temporal graph neural networks fail at three core temporal reasoning skills—counting periodic repetitions, inferring delayed causes, and tracking long-range spatio-temporal dependencies—and that no single…

desk verdict A genuinely useful synthetic benchmark for temporal graph reasoning, with a credible core result but overclaimed conclusions in the long-range task and protocol confounds that need fixing before the claims are settled. read the letter →

arxiv 2507.10183 v2 pith:JECRUHSQ submitted 2025-07-14 cs.LG

Alireza Dizaji , Benedict Aaron Tjandra , Mehrab Hamidi , Shenyang Huang , Guillaume Rabusseau This is my paper
classification cs.LG
keywords temporalgraphlearningdynamiclinkpredictionsyntheticbenchmarkreasoningperiodicitycountingcause-and-effecttaskslong-rangespatio-temporaldependenciesneighborsampling
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

T-GRAB is a set of synthetic dynamic graphs that isolate three temporal skills: counting periodic repetitions, inferring delayed causes, and tracking dependencies that span both space and time. The paper argues that current temporal graph neural networks, across both continuous-time and discrete-time families, do not reliably perform these skills, and that no single method is best at all three. On periodic tasks many models converge on a persistence-like shortcut, copying the previous snapshot rather than counting, while all models degrade on long causal or spatial lags. If right, the benchmark provides a controlled diagnostic that real-world benchmarks cannot, letting researchers attribute failures to specific missing temporal abilities rather than to dataset noise or scale.

What carries the argument

The benchmark is built from three parametric task families. Pdet(k,n) and Psto(k,n) generate deterministic or stochastic periodic graphs where k distinct structures each appear n consecutive times, testing counting and memorization. CE(ℓ) generates random graphs plus a memory node that connects to nodes active ℓ time steps earlier, testing delayed causality. LR(ℓ,d) generates P disjoint paths of length d from a source node and connects a target to the path endpoints from ℓ steps earlier, testing joint spatial and temporal reach. The evaluation uses average F1 over all possible test node pairs, and for periodicity it separately reports change-point F1, the timesteps where the active graph switches, so that persistence shortcuts are exposed.

What would settle it

Run T-GRAB's periodic and long-range tasks while allowing each model its own depth and neighbor-sampling budget, for example d stacked message-passing layers for LR(ℓ,d) and sampling windows covering the full period; if any current model then reaches near-perfect change-point F1 and high F1 at d=16 and ℓ=32, the claim of fundamental, architecture-level temporal reasoning failure is refuted, while no improvement would support it.

Watch

Extended reading notes

Core claim

The central discovery is a benchmark-driven negative result: across the three task families, performance depends sharply on both the task and the architecture, and no model dominates. GC-LSTM handles periodicity best, including the harder stochastic periodicity settings; DyGFormer, TGAT, and TGN degrade most gracefully on delayed causality; and TGAT and TGN hold up best on long-range spatio-temporal tasks, while CTAN, a method explicitly designed for long-range propagation, performs poorly. The paper also shows that many models' high overall scores on periodic tasks come from copying recent snapshots rather than true counting, as clearly revealed by their collapse at change points, where EdgeBank, a non-learning baseline, can outperform them. These patterns point to fundamental shortcomings in current temporal graph learning methods rather than to isolated dataset-specific failures.

Load-bearing premise

The conclusion that current temporal graph neural networks lack these abilities rests on treating the uniform training setup—one message-passing layer, a fixed learning rate and batch size, and 20 sampled neighbors—as fair to every model family, so that observed failures reflect architecture rather than configuration.

Editorial extensions

If this is right

  • High overall scores on temporal link prediction benchmarks can hide a persistence shortcut, so reporting change-point or switch-timing performance is necessary before claiming temporal reasoning ability.
  • Counting periodic structure is not a solved capability: at n=32 EdgeBank can beat all tested temporal graph neural networks, and most models learn to react to a new snapshot rather than anticipate it.
  • Long-range spatio-temporal reasoning degrades for every tested model beyond spatial distance d=8 and large temporal lags, even for attention- and memory-based methods such as TGAT and TGN.
  • The number of sampled temporal neighbors is a high-impact hyperparameter for continuous-time methods, not a nuisance detail; increasing it can produce sizeable gains, so default settings such as 20 neighbors may understate these models' capabilities.
  • Complementary strengths across architectures—recurrent models for periodicity, attention and memory models for causality and long-range dependencies—motivate hybrid designs rather than a single universal temporal graph architecture.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's diagnostic conclusions are conditional on a uniform training protocol: one message-passing layer, a fixed learning rate and batch size, and 20 sampled neighbors for all continuous-time models. Deeper propagation or task-specific tuning could in principle close some of the observed gaps, so the 'fundamental shortcoming' reading should be tested by ablating these settings.
  • The long-range spatio-temporal task likely conflates architectural depth with temporal reasoning, because paths of length d require d spatial hops while the tested models use a single message-passing layer per snapshot; a model with more layers might solve the task without any new temporal machinery.
  • The change-point metric is a transferable idea: any temporal benchmark that includes repeated or periodic structure could report performance at transition timesteps, giving a direct measure of whether a model anticipates change rather than reacts to it.
  • T-GRAB's task families could be composed to test combined skills—for example periodic causes with long delays or periodic long-range paths—which the paper leaves as future work but which would stress models beyond what any single task does alone.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces T-GRAB, a synthetic diagnostic benchmark for temporal graph learning consisting of three task families: periodicity tasks (Pdet, Psto) that test counting and memorization, delayed cause-and-effect tasks (CE(ℓ)) that test temporal memory, and long-range spatio-temporal tasks (LR(ℓ,d)) that test joint spatial and temporal reasoning. The authors evaluate 11 methods (CTDG, DTDG, static GNNs, and simple baselines) under a unified protocol and report that no single model wins all tasks, that many models learn a persistence-like shortcut at periodicity change points, and that performance degrades with larger lags and spatial distances. They also analyze the effect of temporal neighbor sampling for CTDG methods. The central claim is that current TGNNs have fundamental shortcomings in these core temporal reasoning skills.

Significance. If the protocol genuinely isolates the intended temporal skills, T-GRAB would be a valuable contribution in the spirit of bAbI and CLEVR: a controlled, interpretable suite with explicit generation algorithms, exhaustive all-pairs F1 evaluation that avoids negative-sampling bias, open-source code and data, and inclusion of non-neural baselines (EdgeBank, persistence) that provide meaningful lower bounds. The finding that attention/memory CTDG models beat recurrent DTDG models on cause-and-effect while GC-LSTM excels at periodicity is also a useful qualitative result. However, the strength of the paper's conclusions is not yet matched by the evidence: the long-range spatio-temporal task is confounded by a fixed one-layer message-passing architecture, and the cause-and-effect results at large lags are entangled with the fixed neighbor-sampling budget. These issues are local to specific experiments and fixable with additional ablations, so the benchmark's core idea remains promising.

major comments (4)
  1. [§3.3, Appendix D.2, Figure 8] The long-range spatio-temporal task is confounded by the fixed single message-passing layer described in Appendix D.2. In LR(ℓ,d), predicting the target node v_T's connection to the endpoints of length-d paths requires information to propagate along d edges within a snapshot. With one GNN layer, every model has a receptive field of only 1-hop neighbors in the current snapshot, so for d=2,4,8,16 the models are structurally unable to reach the path endpoints regardless of their temporal memory. The monotone drop in F1 as d increases in Figure 8 is therefore at least partly a measure of GNN depth rather than temporal reasoning depth. To support the claim that this task isolates long-range spatio-temporal reasoning, the authors should vary the number of message-passing layers (e.g., 1, 2, 4, 8, 16) or use models with sufficient propagation depth, and report performance as a function of both depth and d.
  2. [§4, Figure 9, Appendix D.2] The fixed temporal neighbor sampling budget of 20 for CTDG methods is too small relative to the lags used in the cause-and-effect tasks (ℓ up to 256). To predict the memory node's edges at time t, a model must retrieve information from t−ℓ, but sampling only the 20 most recent neighbors may exclude the relevant historical interactions entirely. The paper's own Figure 9 shows that performance on CE(64) and CE(256) continues to improve with larger sampling budgets, so the observed degradation at large ℓ in Figure 6 may reflect the sampling protocol rather than an intrinsic memory limit. The authors should rerun the cause-and-effect experiments with neighbor budgets at least as large as the maximum lag (e.g., 256 or 512) and state whether the qualitative ranking changes.
  3. [All experiments, Table 2] All results are reported as point estimates averaged over five seeds, with no error bars, standard deviations, or significance tests. Since the paper's main message is 'no single model consistently outperforms across all tasks' and Table 2 presents fine-grained average ranks, the absence of variance information makes it impossible to know whether adjacent ranks (e.g., TGAT versus TGN in several columns) are meaningfully different. The authors should either include standard deviations in the figures and tables or provide a statistical comparison (e.g., paired tests across seeds).
  4. [Abstract, §5] The abstract and conclusion describe the findings as 'fundamental shortcomings' of TGNNs. This wording is stronger than the evidence can support while the long-range task is depth-limited and the cause-and-effect task is sampling-limited, as noted above. The claims should be scoped to 'shortcomings under the standard evaluation protocol with one message-passing layer and 20 sampled neighbors,' or the additional ablations should be provided to justify the stronger statement.
minor comments (5)
  1. [Table 2] The column header 'Psto(2,n)' appears to be a typo: Section 3.1 defines the stochastic tasks as Psto(k,1), and the results vary k, not n. The header should read 'Psto(k,1)' to match the text and the two probability settings p=0.5 and p=0.9.
  2. [Appendix H.2] The variable k is used to denote the number of sampled neighbors in this appendix, but k already denotes the period length in Section 3.1. This overloaded notation makes the sentences 'increasing k leads to substantial F1 score improvements' confusing; please rename the neighbor count (e.g., m or s).
  3. [Appendix D.2] The sentence 'without enforcing a fixed maximum number of epochs' is slightly awkward; also, the crucial detail that all models use a single message-passing layer appears only in the appendix and should be stated prominently in the main evaluation section, since it materially affects the interpretation of the long-range task.
  4. [Algorithm 1] The pseudocode in Algorithm 1 contains a stray '1 .' line ('1 . F1(best) ← 0;') that appears to be a formatting corruption, and the variable naming is inconsistent (e.g., 'F1(best)' versus 'F(best)1'). Please clean up the pseudocode.
  5. [Figure 3] In the caption and axis labels, the two rows are labeled 'Methods' for both CTDG and DTDG, which is redundant; also, the figure would be more readable if the four panels were individually labeled with the task (Pdet(2,n) all timesteps, Pdet(2,n) change points, Pdet(k,1) all timesteps).

Circularity Check

1 steps flagged · score 6.0 of 10

The long-range spatial finding is entailed by the fixed one-layer message-passing protocol; all other benchmark claims are self-contained.

  1. other [Section 3.3 (Definition 4, long-range spatio-temporal task), Appendix D.2 (method training details), Fig. 8 results paragraph.]
    "We used a single layer of message passing in all of our experiments. (Appendix D.2) Definition 4: "The edge set Et consists of P disjoint paths of length d from the source node vS through randomly chosen intermediate nodes." Results: "all models, even the strongest temporal reasoners, show significant performance degradation when spatial distance exceeds d = 8.""

    The LR task is designed to measure d-hop spatial reasoning, but Appendix D.2 fixes one message-passing layer for every model, capping each model's per-snapshot receptive field at one hop. Node features are constant one-hot IDs, and intermediate/endpoint nodes are drawn without replacement afresh each snapshot, so neither 1-hop neighborhoods nor historical memory can identify which nodes lie at distance d from vS; success requires propagating the vS signal along d edges within the snapshot, which the protocol excludes. The observed monotone F1 collapse with d (Fig. 8) and the conclusion of a persistent challenge of capturing long-range spatial dependencies are therefore consequences of the benchmark's own fixed-depth construction, not an independent measurement of TGNN capability.

full rationale

T-GRAB is an empirical benchmark rather than a derivation, so most circularity patterns do not apply. All three task families are generated by explicit algorithms with ground-truth link structure (Definitions 2-4), and models are scored on held-out snapshots against external baselines (persistence and EdgeBank), so there is no fitted parameter later relabeled as a prediction and no result is constructed to favor a chosen model family; indeed different families win on different tasks (Table 2). Self-citations ([5], [8], [9], [24], all including present authors) are not load-bearing: [9] provides the standard snapshot-to-event translation for running CTDG methods on DTDG data, and the benchmark's validity rests on its own generative processes, not on any cited uniqueness theorem or ansatz. The periodicity tasks properly neutralize the persistence shortcut by reporting change-point F1, and the cause-and-effect task restricts evaluation to the memory node's edges, the only predictable edges by construction of the generative model. The one genuinely by-construction element is the spatial half of the long-range task: Appendix D.2 fixes a single message-passing layer for every model, capping each model's receptive field at one hop per snapshot, while the task (Definition 4) requires recognizing the endpoints of length-d paths from vS, which cannot be done from one-hop features with freshly drawn random paths; the monotone F1 decline with d in Fig. 8 and the long-range-spatial challenge conclusion are entailed by this protocol choice rather than separately measured. A related but openly disclosed confound is the default 20-neighbor sampling window versus lags up to ell=256 in the CE task, which the authors themselves ablate in Section 4 (Fig. 9). Overall, the circularity is partial and limited to one half of one of the three tasks; the counting, memorization, and temporal-memory findings, and the no-single-winner ranking, retain independent empirical content.

Assumptions & free parameters 4 free parameters · 4 assumptions · 2 invented entities

As a benchmark paper, the ledger entries are experimental design choices and synthetic constructs rather than fitted constants or invented physical entities. The most influential hand-chosen settings are the single GNN layer and the default 20-neighbour sampling, both of which directly affect the main findings. The axioms capture the assumptions that the tasks measure what they claim to measure, that the training protocol is fair, that the evaluation metric is reliable, and that the transductive scope is sufficient for the conclusions.

free parameters (4)
  • Number of temporal neighbours sampled (default 20) = 20
    Hand-chosen default for all CTDG methods; Section 4 (Figure 9) shows performance on periodic and cause-effect tasks changes substantially with this setting, so the benchmark conclusions about model capability depend on this choice.
  • Number of GNN message-passing layers = 1
    Appendix D.2 states a single layer of message passing is used in all experiments; the long-range spatio-temporal tasks require propagating information along paths of length d up to 16, so a one-layer GNN cannot, in a single snapshot, aggregate features beyond one hop; this is a critical design choice affecting the claim that models fail on long spatial distances.
  • ER edge probability for cause/effect and periodicity graphs = 0.01
    Controls graph sparsity (about 50 expected edges among 100 nodes); affects task difficulty and the informativeness of F1 scores.
  • Number of paths P in spatio-temporal tasks = 3
    Fixed for all LR(l,d) runs; larger P would increase signal strength for the target node, making the task easier.
assumptions (4)
  • domain assumption Benchmark tasks isolate the targeted temporal capability without exploitable shortcuts
    The paper's diagnostic conclusions assume that high F1 on Pdet(2,n) at change points requires counting, that CE(l) requires retaining exactly l-step history, and that LR(l,d) requires multi-hop spatio-temporal reasoning; the persistence and EdgeBank baselines are used to argue this, but the possibility of other shortcuts (e.g., memorizing full snapshots or using degree/co-occurrence statistics) is not exhaustively excluded.
  • domain assumption Models are trained under a fair, representative protocol so that failures reflect architecture rather than tuning
    All models use the same single-layer architecture, Adam lr 1e-4, batch size one snapshot, and early stopping; no per-model hyperparameter search is performed, and the number of neighbors is varied only in Section 4; performance differences could be artifacts of suboptimal hyperparameters rather than fundamental capability differences.
  • domain assumption Averaged F1 computed over all possible node pairs is an unbiased and reliable diagnostic metric
    The evaluation protocol deliberately avoids negative sampling (Section 3, Evaluation Protocols); this is defensible but assumes that exhaustive F1 over all pairs is well-calibrated for these sparse graphs and that restricting to memory-node edges for CE/LR does not hide model failure modes.
  • domain assumption The transductive setting is sufficient to expose the core temporal reasoning gaps
    The paper's own Limitations section (Appendix A) states that inductive settings and other capabilities (anomaly detection) are left for future work; the conclusions about 'fundamental shortcomings' are scoped to transductive link prediction, and it is assumed these gaps persist in inductive settings.
invented entities (2)
  • Memory node (v_M) in cause-and-effect tasks
    purpose: Acts as the only predictable node linking to nodes active l steps earlier, creating a clean target for measuring delayed causal memory
    The memory node is a synthetic construct in the benchmark generator; it does not correspond to any real-world observed entity and its behaviour is fully specified by the generator, so it provides no out-of-benchmark prediction.
  • Source (v_S) and target (v_T) nodes in spatio-temporal tasks
    purpose: Source seeds P paths of length d each snapshot; target links to path endpoints from l steps earlier, providing a controllable spatio-temporal dependency
    Similarly fully synthetic; part of the dataset construction, not an independently evidenced phenomenon.

how reviews work

0 comments
Cite this review

Pith. "Pith review of T-GRAB: A Synthetic Diagnostic Benchmark for Learning on Temporal Graphs." pith.science (2026). https://pith.science/paper/JECRUHSQ

@misc{pith2026250710183,
  author       = {Pith},
  title        = {Pith review of: T-GRAB: A Synthetic Diagnostic Benchmark for Learning on Temporal Graphs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JECRUHSQ}},
  note         = {Machine review of arXiv:2507.10183}
}
read the original abstract

Dynamic graph learning methods have recently emerged as powerful tools for modelling relational data evolving through time. However, despite extensive benchmarking efforts, it remains unclear whether current Temporal Graph Neural Networks (TGNNs) effectively capture core temporal patterns such as periodicity, cause-and-effect, and long-range dependencies. In this work, we introduce the Temporal Graph Reasoning Benchmark (T-GRAB), a comprehensive set of synthetic tasks designed to systematically probe the capabilities of TGNNs to reason across time. T-GRAB provides controlled, interpretable tasks that isolate key temporal skills: counting/memorizing periodic repetitions, inferring delayed causal effects, and capturing long-range dependencies over both spatial and temporal dimensions. We evaluate 11 temporal graph learning methods on these tasks, revealing fundamental shortcomings in their ability to generalize temporal patterns. Our findings offer actionable insights into the limitations of current models, highlight challenges hidden by traditional real-world benchmarks, and motivate the development of architectures with stronger temporal reasoning abilities. The code for T-GRAB can be found at: https://github.com/alirezadizaji/T-GRAB.

Figures

Figures reproduced from arXiv: 2507.10183 by the authors.

Figure 1
Figure 1. T-GRAB tests the capabilities of TGNNs to reason over time in three fundamental aspects [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Periodic task in P det(k= 2, n= 3) with 2 unique snapshots repeated 3 times within a period. Definition 2 (Periodicity Tasks). Let k, n ∈ N. The task families P det(k, n) and P sto(k, n) are defined based on a repeating pattern where integers i = 1, . . . , k each appears n consecutive times before cycling. For each t, let it = (⌊t/n⌋ mod k) + 1. In P det(k, n), the dynamic graph G = G1, G2, . . . is a periodic sequ… view at source ↗
Figure 3
Figure 3. Performance of CTDG methods (top row) and DTDG methods (bottom row) on the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Performance of TGL methods on the stochastic periodicity tasks [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Illustration of a delayed cause-and-effect dataset with lag [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Methods’ performance on cause-and￾effect tasks CE(ℓ) across five temporal lags ℓ. Results [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Illustration of a long-range spatio-temporal task [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Performance varying spatial and temporal distances in the long-range spatio-temporal tasks. [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Effect of the number of neighbors on periodic (left), and cause-and-effect (right) tasks . [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 10
Figure 10. Figure 10: The F1 score at each timestep over the evaluation period is shown for DTDG (left) and CTDG (right) methods on the P det(2, 32) dataset [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]
Figure 11
Figure 11. Figure 11: The impact of number of historical neighbor sampling on stochastic and deterministic [PITH_FULL_IMAGE:figures/full_fig_p018_11.png]
Figure 12
Figure 12. Figure 12: (Full methods version) Long-range spatio-temporal tasks across varying spatial and [PITH_FULL_IMAGE:figures/full_fig_p019_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 16 canonical work pages

  1. [1]

    Weisfeiler-Lehman goes Dynamic: An Analysis of the Expressive Power of Graph Neural Networks for Attributed and Dynamic Graphs

    Silvia Beddar-Wiesing, Giuseppe Alessio D’Inverno, Caterina Graziani, Veronica Lachi, Alice Moallemy-Oureh, Franco Scarselli, and Josephine Maria Thomas. Weisfeiler-lehman goes dynamic: An analysis of the expressive power of graph neural networks for attributed and dynamic graphs, 2024. URL https://arxiv.org/abs/2210.03990

    work page Pith review arXiv 2024

  2. [2]

    Gc-lstm: Graph convolution embedded lstm for dynamic link prediction, 2021

    Jinyin Chen, Xueke Wang, and Xuanheng Xu. Gc-lstm: Graph convolution embedded lstm for dynamic link prediction, 2021. URL https://arxiv.org/abs/1812.04206

    arXiv 2021

  3. [3]

    On the evolution of random graphs

    Paul Erdos and Alfréd Rényi. On the evolution of random graphs. Publ. math. inst. hung. acad. sci, 5(1):17–60, 1960

    work page 1960

  4. [4]

    A comprehensive survey of dynamic graph neural networks: Models, frameworks, benchmarks, experiments and challenges, 2024

    ZhengZhao Feng, Rui Wang, TianXing Wang, Mingli Song, Sai Wu, and Shuibing He. A comprehensive survey of dynamic graph neural networks: Models, frameworks, benchmarks, experiments and challenges, 2024. URL https://arxiv.org/abs/2405.00476

    arXiv 2024

  5. [5]

    Tgb 2.0: A benchmark for learning on temporal knowledge graphs and heterogeneous graphs, 2024

    Julia Gastinger, Shenyang Huang, Mikhail Galkin, Erfan Loghmani, Ali Parviz, Farimah Pour- safaei, Jacob Danovitch, Emanuele Rossi, Ioannis Koutis, Heiner Stuckenschmidt, Reihaneh Rabbany, and Guillaume Rabusseau. Tgb 2.0: A benchmark for learning on temporal knowledge graphs and heterogeneous graphs, 2024. URL https://arxiv.org/abs/2406.09639

    arXiv 2024

  6. [6]

    Long range propagation on continuous-time dynamic graphs, 2024

    Alessio Gravina, Giulio Lovisotto, Claudio Gallicchio, Davide Bacciu, and Claas Grohnfeldt. Long range propagation on continuous-time dynamic graphs, 2024. URL https://arxiv. org/abs/2406.02740

    arXiv 2024

  7. [7]

    Stochastic blockmodels: First steps

    Paul W Holland, Kathryn Blackmond Laskey, and Samuel Leinhardt. Stochastic blockmodels: First steps. Social networks, 5(2):109–137, 1983

    1983

  8. [8]

    Tempo- ral graph benchmark for machine learning on temporal graphs

    Shenyang Huang, Farimah Poursafaei, Jacob Danovitch, Matthias Fey, Weihua Hu, Emanuele Rossi, Jure Leskovec, Michael Bronstein, Guillaume Rabusseau, and Reihaneh Rabbany. Tempo- ral graph benchmark for machine learning on temporal graphs. Advances in Neural Information Processing Systems, 36:2056–2073, 2023

    work page 2023

Show all 34 references
  1. [9]

    Utg: Towards a unified view of snapshot and event based models for temporal graphs,

    Shenyang Huang, Farimah Poursafaei, Reihaneh Rabbany, Guillaume Rabusseau, and Emanuele Rossi. Utg: Towards a unified view of snapshot and event based models for temporal graphs,

  2. [10]

    Lawrence Zitnick, and Ross Girshick

    Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C. Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning, 2016. URL https://arxiv.org/abs/1612.06890

    2016 arXiv

  3. [11]

    Representation learning for dynamic graphs: A survey

    Seyed Mehran Kazemi, Rishab Goel, Kshitij Jain, Ivan Kobyzev, Akshay Sethi, Peter Forsyth, and Pascal Poupart. Representation learning for dynamic graphs: A survey. Journal of Machine Learning Research, 21(70):1–73, 2020

    2020

  4. [12]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks, 2017. URL https://arxiv.org/abs/1609.02907

    2017 arXiv

  5. [13]

    Predicting dynamic embedding trajectory in temporal interaction networks

    Srijan Kumar, Xikun Zhang, and Jure Leskovec. Predicting dynamic embedding trajectory in temporal interaction networks. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , KDD ’19, page 1269–1278. ACM, July 2019. doi: 10.1145/3...

    2019

  6. [14]

    Neighborhood-aware scalable temporal network representation learning

    Yuhong Luo and Pan Li. Neighborhood-aware scalable temporal network representation learning. In Learning on Graphs Conference, pages 1–1. PMLR, 2022

    2022

  7. [15]

    Sandy" Pentland. Sensing the

    Anmol Madan, Manuel Cebrian, Sai Moturu, Katayoun Farrahi, and Alex "Sandy" Pentland. Sensing the "health state" of a community. IEEE Pervasive Computing, 11(4):36–45, 2012. doi: 10.1109/MPRV .2011.79. 10

    2012 doi

  8. [16]

    Behaviour suite for reinforcement learning, 2020

    Ian Osband, Yotam Doron, Matteo Hessel, John Aslanides, Eren Sezener, Andre Saraiva, Katrina McKinney, Tor Lattimore, Csaba Szepesvari, Satinder Singh, Benjamin Van Roy, Richard Sutton, David Silver, and Hado Van Hasselt. Behaviour suite for reinforcement learning, 2020. URL h...

    2020 arXiv

  9. [17]

    Pietro Panzarasa, Tore Opsahl, and Kathleen M. Carley. Patterns and dynamics of users’ behavior and interaction: Network analysis of an online community. J. Am. Soc. Inf. Sci. Technol., 60(5):911–932, May 2009. ISSN 1532-2882

    2009

  10. [18]

    Schardl, and Charles E

    Aldo Pareja, Giacomo Domeniconi, Jie Chen, Tengfei Ma, Toyotaro Suzumura, Hiroki Kaneza- shi, Tim Kaler, Tao B. Schardl, and Charles E. Leiserson. Evolvegcn: Evolving graph convolu- tional networks for dynamic graphs, 2019. URL https://arxiv.org/abs/1902.10191

    2019 arXiv

  11. [19]

    Towards better evaluation for dynamic link prediction, 2022

    Farimah Poursafaei, Shenyang Huang, Kellin Pelrine, and Reihaneh Rabbany. Towards better evaluation for dynamic link prediction, 2022. URL https://arxiv.org/abs/2207.10128

    2022 arXiv

  12. [20]

    Temporal graph networks for deep learning on dynamic graphs, 2020

    Emanuele Rossi, Ben Chamberlain, Fabrizio Frasca, Davide Eynard, Federico Monti, and Michael Bronstein. Temporal graph networks for deep learning on dynamic graphs, 2020. URL https://arxiv.org/abs/2006.10637

    2020 arXiv

  13. [21]

    Graphpulse: Topological representations for temporal graph property prediction

    Kiarash Shamsi, Farimah Poursafaei, Shenyang Huang, Bao Tran Gia Ngo, Baris Coskunuzer, and Cuneyt Gurcan Akcora. Graphpulse: Topological representations for temporal graph property prediction. In The Twelfth International Conference on Learning Representations, 2024

    2024

  14. [22]

    The enron email dataset database schema and brief statistical report

    Jitesh Shetty and Jafar Adibi. The enron email dataset database schema and brief statistical report. 2004. URL https://api.semanticscholar.org/CorpusID:59919272

    2004

  15. [23]

    Static graph approximations of dynamic contact networks for epidemic forecasting

    Razieh Shirzadkhani, Shenyang Huang, Abby Leung, and Reihaneh Rabbany. Static graph approximations of dynamic contact networks for epidemic forecasting. Scientific Reports, 14 (1):11696, 2024

    2024

  16. [24]

    Towards neural scaling laws for foundation models on temporal graphs

    Razieh Shirzadkhani, Tran Gia Bao Ngo, Kiarash Shamsi, Shenyang Huang, Farimah Poursafaei, Poupak Azad, Reihaneh Rabbany, Baris Coskunuzer, Guillaume Rabusseau, and Cuneyt Gurcan Akcora. Towards neural scaling laws for foundation models on temporal graphs. arXiv preprint arXiv...

    2024 arXiv

  17. [25]

    Foundations and modeling of dynamic networks using dynamic graph neural networks: A survey

    Joakim Skarding, Bogdan Gabrys, and Katarzyna Musial. Foundations and modeling of dynamic networks using dynamic graph neural networks: A survey. IEEE Access, 9:79143–79168, 2021. ISSN 2169-3536. doi: 10.1109/access.2021.3082932. URL http://dx.doi.org/10.1109/ ACCESS.2021.3082932

    2021

  18. [26]

    Souza, Diego Mesquita, Samuel Kaski, and Vikas Garg

    Amauri H. Souza, Diego Mesquita, Samuel Kaski, and Vikas Garg. Provably expressive temporal graph networks, 2022. URL https://arxiv.org/abs/2209.15059

    2022 arXiv

  19. [27]

    Enhancing the expressivity of temporal graph networks through source-target identification, 2024

    Benedict Aaron Tjandra, Federico Barbero, and Michael Bronstein. Enhancing the expressivity of temporal graph networks through source-target identification, 2024. URL https://arxiv. org/abs/2411.03596

    2024 arXiv

  20. [28]

    Graph attention networks, 2018

    Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. Graph attention networks, 2018. URL https://arxiv.org/abs/1710.10903

    2018 arXiv

  21. [29]

    Rush, Bart van Merriënboer, Armand Joulin, and Tomas Mikolov

    Jason Weston, Antoine Bordes, Sumit Chopra, Alexander M. Rush, Bart van Merriënboer, Armand Joulin, and Tomas Mikolov. Towards ai-complete question answering: A set of prerequisite toy tasks, 2015. URL https://arxiv.org/abs/1502.05698

    2015 arXiv

  22. [30]

    Inductive representation learning on temporal graphs, 2020

    Da Xu, Chuanwei Ruan, Evren Korpeoglu, Sushant Kumar, and Kannan Achan. Inductive representation learning on temporal graphs, 2020. URL https://arxiv.org/abs/2002. 07962

    2020

  23. [31]

    Tgb-seq benchmark: Challenging temporal gnns with complex sequential dynamics, 2025

    Lu Yi, Jie Peng, Yanping Zheng, Fengran Mo, Zhewei Wei, Yuhang Ye, Yue Zixuan, and Zengfeng Huang. Tgb-seq benchmark: Challenging temporal gnns with complex sequential dynamics, 2025. URL https://arxiv.org/abs/2502.02975. 11

    2025 arXiv

  24. [32]

    Towards better dynamic graph learning: New architecture and unified library, 2023

    Le Yu, Leilei Sun, Bowen Du, and Weifeng Lv. Towards better dynamic graph learning: New architecture and unified library, 2023. URL https://arxiv.org/abs/2303.13047

    2023 arXiv

  25. [33]

    T-gcn: A temporal graph convolutional network for traffic prediction

    Ling Zhao, Yujiao Song, Chao Zhang, Yu Liu, Pu Wang, Tao Lin, Min Deng, and Haifeng Li. T-gcn: A temporal graph convolutional network for traffic prediction. IEEE Transactions on Intelligent Transportation Systems, 21(9):3848–3858, September 2020. ISSN 1558-0016. doi: 10.1109/...

    2020

  26. [2024]

    URL https://arxiv.org/abs/2407.12269

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.