Pith. sign in

REVIEW 3 major objections 6 minor 46 references

TIDFormer: Exploiting Temporal and Interactive Dynamics Makes A Great Dynamic Graph Transformer

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read TIDFormer claims that defining self-attention over historical interactions, instead of single nodes or mixed node sequences, is what makes a dynamic graph Transformer accurate and efficient.

desk verdict Solid empirical architecture paper, but the BIE module's batch-dependent retrieval could break causality and needs a straight answer before the SOTA claims carry weight. read the letter →

arxiv 2506.00431 v2 pith:LOTV6FSF submitted 2025-05-31 cs.LG

Jie Peng , Zhewei Wei , Yuhang Ye This is my paper
classification cs.LG
keywords continuous-timedynamicgraphslinkpredictiongraphtransformerinteraction-levelself-attentiontemporalencodingneighborinteractiontime-seriesdecompositionbipartite
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

TIDFormer is a Transformer for continuous-time dynamic graphs that claims to settle how self-attention should be defined on such graphs: at the interaction level, where each token is a historical interaction between a source and a target node and the original chronological order is preserved. The paper argues that earlier single-node and multi-node tokenizations either ignore node-pair information or mix independent timelines, and that this explains their weaker link prediction. On top of this interaction-level tokenization, TIDFormer adds three simple encoders—calendar-based mixed-granularity time encoding, bidirectional first-order neighbor interaction encoding, and a trend/seasonality decomposition—and reports state-of-the-art average precision over seven datasets under transductive and inductive settings with three negative sampling strategies. The paper further claims these gains come at lower cost than previous Transformer-based dynamic graph models, because the interaction encoding obtains higher-order information by dictionary lookup rather than by explicit higher-order sampling.

What carries the argument

The load-bearing object is the interaction-level self-attention (SAM at IL): a sequence $H_{IL}\in\mathbb{R}^{n\times(d+d_i)}$ in which each token combines the node and edge features of one historical interaction with a temporal-interactive embedding $I$, preserving the original chronological order. That embedding is the concatenation of three cheap encoders: Mixed-granularity Temporal Encoding (MTE), which adds coarse calendar segment information $\theta\Delta T^{coarse}$ to the fine-grained cosine encoding $\cos(\omega\Delta T^{fine})$; Bidirectional Interaction Encoding (BIE), which for each first-order neighbor not shared by both endpoints retrieves that neighbor's first-order sequence from batch-level source/target dictionaries and splices it in, giving second-order reach without second-order sampling; and Seasonality & Trend Encoding (STE), which applies average pooling to the interaction sequence to obtain trend $Z_t$ and seasonal residual $Z_s = Q - Z_t$. These modules feed the multi-head self-attention in a standard Transformer block, and the paper's efficiency argument is that MTE and STE are $O(n)$, BIE is $O(n)$ via constant-time hash lookups, leaving the $O(n^2 \times dim)$ Transformer as the dominant term.

What would settle it

Run the same TIDFormer configuration on a CTDG dataset with different batch sizes and with batches whose edges are shuffled or drawn from disjoint time windows, and compare AP against DyGFormer and FreeDyG; if the advantage shrinks or reverses when the batch dictionary contents change, the BIE lookup—not the interaction-level attention—is carrying the result.

Watch

Extended reading notes

Core claim

The paper's central claim is that the key to a successful dynamic graph Transformer is to tokenize each historical event as an interaction between a source and a target node and to compute attention over those interaction tokens in their true temporal order. This interaction-level self-attention is said to be interpretable: the attention weights it learns are correlated with node interaction frequency, so influential key nodes receive higher attention, whereas the multi-node mixed-sequence attention of prior work does not show this correlation. The paper also claims that two lightweight encoders—Bidirectional Interaction Encoding, which reconstructs source and target sequences from first-order neighbor dictionaries maintained over the current batch, and Seasonality & Trend Encoding, which decomposes the interaction sequence into trend and seasonal parts by average pooling—jointly capture interactive and temporal dynamics without extra complex modules. Together with Mixed-granularity Temporal Encoding, which adds calendar-based weekly/monthly/yearly segments to fine-grained cosine time embeddings, the full model is reported to rank first on most of seven CTDG benchmarks across transductive and inductive link prediction and to be faster per epoch than previous Transformer baselines.

Load-bearing premise

The load-bearing premise is that the BIE module's batch-level dictionaries give a stable, unbiased estimate of each node's true higher-order neighborhood; if the composition of the batch changes what information a node can retrieve, then the model's per-edge representation and the reported gains may be properties of the batching protocol rather than of the interaction-level design.

Editorial extensions

If this is right

  • For bipartite graphs, where source and target nodes share no common first-order neighbors, the bidirectional dictionary reconstruction gives BIE a way to carry higher-order interaction information, so interaction encoding no longer degenerates to counting historical interaction times.
  • An interpretable attention definition on dynamic graphs means the attention weights themselves can be inspected: nodes with high interaction frequency should receive proportionally higher weights, giving practitioners a diagnostic signal.
  • The total time complexity stays $O(n) + O(n^2 \times dim)$, the same order as prior first-order Transformer methods, while avoiding the $O(n^2)$ second-order sampling used by stacked TGAT/TGN-style models.
  • If the reported average ranks hold, TIDFormer is a default choice for continuous-time dynamic link prediction across social, interaction, and proximity networks.

Reading between the lines

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

  • Editorial inference: the BIE module's dictionaries are built from the current batch, so the method's accuracy could depend on batch composition; a natural test is to vary batch size or shuffle batch membership and check whether the reported AP advantage persists.
  • Editorial inference: the calendar segmentation in MTE is chosen by dataset duration; the same additive coarse-grained term could be replaced by a learnable partition or by adaptive period detection, which might remove the need to pick weekly/monthly/yearly per dataset.
  • Editorial inference: the trend-season decomposition is generic enough to be reused as a drop-in input encoding in other sequence-based temporal graph models, not only Transformers.
  • Editorial inference: the reported correlation between attention weights and interaction frequency suggests a cheap interpretability check for any future dynamic graph Transformer, independent of link prediction accuracy.
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

3 major / 6 minor

Summary. The paper proposes TIDFormer, a continuous-time dynamic graph Transformer whose central design is an interaction-level self-attention mechanism. Each historical interaction is tokenized by combining node and edge features with a temporal-interactive embedding built from three modules: Mixed-granularity Temporal Encoding (MTE), Bidirectional Interaction Encoding (BIE), and Seasonality & Trend Encoding (STE). The authors argue that this SAM is interpretable, that BIE enlarges receptive fields using only first-order neighbor sampling, and that TIDFormer outperforms state-of-the-art baselines on seven dynamic graph datasets under transductive and inductive settings with three negative sampling strategies, while being more efficient than prior Transformer-based methods.

Significance. If the empirical claims are valid, the paper makes a useful contribution: the interaction-level SAM is a principled tokenization for link prediction, the calendar-based MTE is simple and plausible, and BIE is an interesting way to incorporate higher-order interaction information without explicit higher-order sampling. The evaluation is extensive, covering seven datasets, three negative sampling strategies, and transductive/inductive splits, and the authors state that source code is released. The main obstacle is that the BIE module as described may use future information within a batch, which would undermine the reported gains; the interpretability and SOTA claims also need stronger quantitative support.

major comments (3)
  1. [Section 4.3, Eq. (8)] The BIE module retrieves second-order neighbor sequences from src-Dict and tgt-Dict, which are described as storing the m source and target sequences in the current training or test batch. The paper never states that these dictionary entries are restricted to interactions with timestamps strictly before the timestamp of the edge being processed. Under the chronological train/validation/test splits described in Section 5.1, a test batch contains edges with different timestamps; for an edge (s,t,t0), the dictionary entry for a neighbor v may be built from a batch edge (u,v,t1) with t1 > t0. The BIE token for (s,t,t0) would then encode interactions after t0, violating the causal chronological order that the interaction-level SAM is claimed to preserve. Since the ablation in Figure 4 shows that removing BIE degrades performance, the strong average ranks in Tables 1 and 5 could be partly an artifact of the batch evaluation protocol rather than of the learned per-edge dynamics. Please specify a causal construction, for example processing each batch in timestamp order and updating the dictionaries only with edges whose timestamp is strictly smaller than the current edge's timestamp, or rerun the experiments under such a constraint and report whether the conclusions change.
  2. [Section 4.2, Figure 2] The interpretability verification is qualitative and limited. The claim of a 'strong correlation' between node interaction frequency and learned attention weights is based on visual inspection of one dataset (MOOC) at two training snapshots, with no correlation coefficient, confidence interval, or comparison across datasets. Please report quantitative measures such as Spearman rank correlation for all three SAM types, ideally across several datasets, and include error bars over seeds.
  3. [Section 5.2, Tables 1 and 5] The headline claim that TIDFormer 'outperforms SOTA models across most datasets' is not fully supported by the reported numbers. For example, in Table 1 under the transductive/random setting, TIDFormer's Wikipedia AP (99.30±0.03) is below FreeDyG and TPNet (99.32), and on Reddit it is below FreeDyG (99.38 vs 99.48). Many differences from the runner-up are within one standard deviation. Please add significance tests, or at least explicitly discuss overlapping confidence intervals, and qualify the claims accordingly.
minor comments (6)
  1. [Section 4.3, Eq. (7)] The notation 'or' leaves two alternative MTE formulations; please specify which one is used in the experiments and whether it is addition or concatenation.
  2. [Section 4.3, Eq. (8)] The notation HH_{tgt2→{src2,src4}} is undefined; please define the reconstruction operation and the interaction-count matrix I clearly.
  3. [Section 4.3, STE] The input Q is described as the 'normalized index of the interacted neighbor nodes'; the normalization is not defined, and it is unclear why average pooling and subtraction on normalized indices produce meaningful trend and seasonal components. Please justify this choice or replace it with a decomposition on learned embeddings.
  4. [Section 5.4, Figure 5] The efficiency comparison is run on a single dataset and includes only TGAT, TGN, and DyGFormer; it does not cover FreeDyG, RepeatMixer, or TPNet, which are central to the effectiveness tables, so the efficiency claim should be scoped accordingly.
  5. [Section 5.1, Appendix B.1] The batch size (200) is a key parameter for BIE because the dictionaries are built from the current batch; please report sensitivity of the results to batch size and confirm that the reported numbers are stable.
  6. [Section 4 heading] There is a typo in the section title: 'Metholodgy' should be 'Methodology'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: TIDFormer's central claims are held-out test-set evaluations against external baselines, and the proposed modules are specified by explicit equations rather than fitted to the target results.

full rationale

TIDFormer's load-bearing assertion is empirical: it reports link prediction AP and node classification AUC-ROC on seven continuous-time dynamic graph datasets using the standard 70%/15%/15% chronological split and three negative sampling strategies, comparing against ten external baselines. These are genuine test-set predictions, not fitted parameters renamed as predictions. The MTE, BIE, and STE modules are each defined by explicit formulas (Equations 6-10), and no module is fitted to the test labels or to the reported performance metric. The BIE module does retrieve second-order neighbor sequences from dictionaries built from the current batch, which raises a legitimate correctness and information-leakage concern if batches are not time-ordered, but this is an evaluation-protocol risk rather than circularity: the predictions are still model outputs on held-out edges, and no equation is equivalent to its own input by construction. The interpretability analysis is a post-hoc diagnostic correlating learned attention weights with interaction frequency; although the BIE tokenization supplies frequency-related features, the paper does not claim to derive a theorem from this correlation, and the diagnostic does not reduce a prediction to a fitted input. Self-citations (references [4], [16], [18], [37], and [44]) are confined to background and related-work discussion and are not load-bearing for the proposed architecture's correctness or for the empirical comparison. No uniqueness theorem, ansatz, or known result is imported from the authors' own prior work to force the design choice. Therefore the derivation chain is self-contained with respect to the reported experiments, and no circular step can be exhibited.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim rests on standard ML components plus several hand-chosen hyperparameters (neighbor count, calendar segment R, alpha/beta) and domain assumptions (first-order sampling sufficiency, calendar relevance, batch-dictionary validity, frequency-as-importance). The most fragile are the batch-dependent BIE and the ad hoc STE.

free parameters (5)
  • number_of_sampled_first_order_neighbors = per dataset, e.g., 16 for MOOC, larger for Social Evo
    Searched over [16, 24, 32, 40, 48, 56, 64] (Table 4); the chosen value varies by dataset and affects the sequence length, hence the attention complexity and performance.
  • calendar_segment_R_and_granularity = e.g., weekly&4, yearly&5, monthly&8, weekly&28
    Chosen by hand from each dataset's duration (Table 3 and Section B.1); R determines the coarse-grained time embedding magnitude theta = {1/R}^{dt} in MTE (Eq. 6-7).
  • MTE_alpha_beta = not reported
    Hyperparameters alpha, beta in omega = {alpha^{-(j-1)/beta}} are said to be selected so that DeltaT_max * alpha^{-(j-1)/beta} tends to 0; values are not given.
  • transformer_hidden_dim_and_heads = 2 layers, 2 heads, hidden dim searched
    Standard hyperparameters (Section B.1), tuned via grid search on the random NSS validation set.
  • batch_size = 200
    Batch size m is used for BIE dictionaries; larger m provides more second-order retrievals, so it is a free parameter affecting the BIE encoding.
assumptions (5)
  • domain assumption First-order neighbor sampling with a fixed-length chronological window is a sufficient representation of a node's temporal neighborhood.
    The model builds all sequences from n sampled first-order neighbors (Section 4.3), a standard assumption in DGNNs (DyGFormer, FreeDyG), but it limits the receptive field.
  • domain assumption Calendar-based coarse time segments (weekly/monthly/yearly) carry signal for dynamic graph link prediction.
    MTE encodes integer bucket indices from calendar partitions (Eq. 6-7); the paper asserts these are informative but does not provide a formal justification.
  • domain assumption In-batch dictionaries (src-Dict, tgt-Dict) give a valid approximation of second-order neighbors without temporal or label leakage.
    BIE retrieves the neighbor sequence of any node in Psi minus Phi from the current batch's dictionaries (Section 4.3, Eq. 8); the paper does not analyze batch-dependence or leakage.
  • domain assumption Interaction frequency of a node is a valid proxy for its importance, so correlation of attention weights with frequency demonstrates interpretability.
    The interpretability analysis (Section 4.2) treats this correlation as evidence; this is a heuristic, not a formal notion of interpretability.
  • ad hoc to paper Average pooling and subtraction on a sequence of normalized neighbor node indices yields meaningful trend and seasonal components.
    STE (Eq. 10) applies time-series decomposition to a categorical node-ID sequence; subtracting a moving average of unordered IDs is not a standard or justified operation for categorical data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TIDFormer: Exploiting Temporal and Interactive Dynamics Makes A Great Dynamic Graph Transformer." pith.science (2026). https://pith.science/paper/LOTV6FSF

@misc{pith2026250600431,
  author       = {Pith},
  title        = {Pith review of: TIDFormer: Exploiting Temporal and Interactive Dynamics Makes A Great Dynamic Graph Transformer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LOTV6FSF}},
  note         = {Machine review of arXiv:2506.00431}
}
read the original abstract

Due to the proficiency of self-attention mechanisms (SAMs) in capturing dependencies in sequence modeling, several existing dynamic graph neural networks (DGNNs) utilize Transformer architectures with various encoding designs to capture sequential evolutions of dynamic graphs. However, the effectiveness and efficiency of these Transformer-based DGNNs vary significantly, highlighting the importance of properly defining the SAM on dynamic graphs and comprehensively encoding temporal and interactive dynamics without extra complex modules. In this work, we propose TIDFormer, a dynamic graph TransFormer that fully exploits Temporal and Interactive Dynamics in an efficient manner. We clarify and verify the interpretability of our proposed SAM, addressing the open problem of its uninterpretable definitions on dynamic graphs in previous works. To model the temporal and interactive dynamics, respectively, we utilize the calendar-based time partitioning information and extract informative interaction embeddings for both bipartite and non-bipartite graphs using merely the sampled first-order neighbors. In addition, we jointly model temporal and interactive features by capturing potential changes in historical interaction patterns through a simple decomposition. We conduct extensive experiments on several dynamic graph datasets to verify the effectiveness and efficiency of TIDFormer. The experimental results demonstrate that TIDFormer excels, outperforming state-of-the-art models across most datasets and experimental settings. Furthermore, TIDFormer exhibits significant efficiency advantages compared to previous Transformer-based methods.

Figures

Figures reproduced from arXiv: 2506.00431 by the authors.

Figure 1
Figure 1. Illustration of three types of SAMs on dynamic graphs at different levels. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the key nodes’ attention weights of three types of SAMs from the start to the end of training on MOOC. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Initial Node and Edge Encoding. Based on the sampled 𝑛 first￾order neighbor nodes, we directly retrieve the neighbor node em￾beddings N ∈ R 𝑛×𝑑𝑛 and the corresponding edge embeddings E ∈ R 𝑛×𝑑𝑒 to construct the initial sequence H = [N, E] ∈ R 𝑛× (𝑑𝑛+𝑑𝑒 ) , with zero-padding guaranteeing the length of the sequence. Mixed-granularity Temporal Encoding (MTE). When modeling real-world sequential data with temporal featu… view at source ↗
Figures from the paper (6 more)
Figure 3
Figure 3. Figure 3: The pipeline of TIDFormer. on 𝑄 𝑠𝑟𝑐0 and 𝑄 𝑡𝑔𝑡0 to reconstruct them. For simplicity, we denote 𝑄 𝑠𝑟𝑐0 ∪ 𝑄 𝑡𝑔𝑡0 as Ψ and 𝑄 𝑠𝑟𝑐0 ∩ 𝑄 𝑡𝑔𝑡0 as Φ. Formally, for each node of 𝑄 𝑠𝑟𝑐0 that is not in Φ, we retrieve its second-order neighbor sequence from tgt-Dict of the current…
Figure 4
Figure 4. Figure 4: The results of the ablation experiments on MOOC. Orange: Comparison of three types of SAMs. Yellow: Performance [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: The AP (%) and training time (s) per epoch in [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: The results of the ablation experiments on LastFM. [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: The results of the ablation experiments on UCI. The color correspondences for the experimental settings are the same [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: The results of the hyperparameter study. [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 27 canonical work pages

  1. [1]

    Unai Alvarez-Rodriguez, Federico Battiston, Guilherme Ferraz de Arruda, Yamir Moreno, Matjaž Perc, and Vito Latora. 2021. Evolutionary dynamics of higher- order interactions in social networks.Nature Human Behaviour5, 5 (2021), 586–595

    2021

  2. [2]

    Oliver D Anderson. 1976. Time-Series. 2nd edn

    work page 1976

  3. [3]

    Khac-Hoai Nam Bui, Jiho Cho, and Hongsuk Yi. 2022. Spatial-temporal graph neural network for traffic forecasting: An overview and open research issues. Applied Intelligence52, 3 (2022), 2763–2774

    work page 2022

  4. [4]

    Ming Chen, Zhewei Wei, Zengfeng Huang, Bolin Ding, and Yaliang Li. 2020. Simple and deep graph convolutional networks. InInternational conference on machine learning. PMLR, 1725–1735

    2020

  5. [5]

    STL Cleveland. [n. d.]. A seasonal-trend decomposition procedure based on Loess (with discussion).J. Off. Stat6, 3 ([n. d.])

    work page

  6. [6]

    Weilin Cong, Yanhong Wu, Yuandong Tian, Mengting Gu, Yinglong Xia, Mehrdad Mahdavi, and Chun-cheng Jason Chen. 2021. Dynamic Graph Representation Learning via Graph Transformer Networks.CoRRabs/2111.10447 (2021)

    arXiv 2021

  7. [7]

    Weilin Cong, Si Zhang, Jian Kang, Baichuan Yuan, Hao Wu, Xin Zhou, Hanghang Tong, and Mehrdad Mahdavi. 2023. Do We Really Need Complicated Model Architectures For Temporal Networks?. InInternational Conference on Learning Representations

    work page 2023

  8. [8]

    Chen Gao, Xiang Wang, Xiangnan He, and Yong Li. 2022. Graph neural net- works for recommender system. InProceedings of the Fifteenth ACM International Conference on Web Search and Data Mining. 1623–1625

    work page 2022

Show all 46 references
  1. [9]

    Zhen Han, Jindong Jiang, Yuyi Wang, Yunpu Ma, and Volker Tresp. 2019. The graph hawkes network for reasoning on temporal knowledge graphs. InLearning with Temporal Point Processes Workshop at at the 33rd Conference on Neural Information Processing Systems (NeurIPS 2019) NeurIPS 2019

    2019

  2. [10]

    Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory.Neural computation9, 8 (1997), 1735–1780

    1997

  3. [11]

    Seyed Mehran Kazemi, Rishab Goel, Kshitij Jain, Ivan Kobyzev, Akshay Sethi, Peter Forsyth, and Pascal Poupart. 2020. Representation Learning for Dynamic Graphs: A Survey.J. Mach. Learn. Res.21 (2020), 70:1–70:73

    2020

  4. [12]

    Thomas N Kipf and Max Welling. 2016. Semi-supervised classification with graph convolutional networks.arXiv preprint arXiv:1609.02907(2016)

    2016 arXiv

  5. [13]

    Srijan Kumar, Xikun Zhang, and Jure Leskovec. 2019. Predicting Dynamic Em- bedding Trajectory in Temporal Interaction Networks. InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. ACM, 1269–1278

    2019

  6. [14]

    Yiming Li, Yanyan Shen, Lei Chen, and Mingxuan Yuan. 2023. Zebra: When tem- poral graph neural networks meet temporal personalized PageRank.Proceedings of the VLDB Endowment16, 6 (2023), 1332–1345

    2023

  7. [15]

    Xiaodong Lu, Leilei Sun, Tongyu Zhu, and Weifeng Lv. 2024. Improving Tem- poral Link Prediction via Temporal Walk Matrix Projection.arXiv preprint arXiv:2410.04013(2024)

    2024 arXiv

  8. [16]

    Jiahong Ma, Mingguo He, and Zhewei Wei. 2023. PolyFormer: Scalable Graph Transformer via Polynomial Attention. (2023)

    2023

  9. [17]

    Franco Manessi, Alessandro Rozza, and Mario Manzo. 2020. Dynamic graph convolutional networks.Pattern Recognition97 (2020), 107000

    2020

  10. [18]

    Jie Peng, Runlin Lei, and Zhewei Wei. 2024. Beyond over-smoothing: Uncovering the trainability challenges in deep graph neural networks. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management. 1878–1887

    2024

  11. [19]

    Farimah Poursafaei, Andy Huang, Kellin Pelrine, and Reihaneh Rabbany. 2022. To- wards Better Evaluation for Dynamic Link Prediction. InThirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track

    2022

  12. [20]

    Emanuele Rossi, Ben Chamberlain, Fabrizio Frasca, Davide Eynard, Federico Monti, and Michael Bronstein. 2020. Temporal Graph Networks for Deep Learning on Dynamic Graphs. InICML 2020 Workshop on Graph Representation Learning

    2020

  13. [21]

    Aravind Sankar, Yanhong Wu, Liang Gou, Wei Zhang, and Hao Yang. 2020. DySAT: Deep Neural Representation Learning on Dynamic Graphs via Self- Attention Networks. InThe Thirteenth ACM International Conference on Web Search and Data Mining. ACM, 519–527

    2020

  14. [22]

    Amit Sharma, Ashutosh Sharma, Polina Nikashina, Vadim Gavrilenko, Alexey Tselykh, Alexander Bozhenyuk, Mehedi Masud, and Hossam Meshref. 2023. A graph neural network (GNN)-based approach for real-time estimation of traffic speed in sustainable smart cities.Sustainability15, 15...

    2023

  15. [23]

    1995.Fourier transforms

    Ian Naismith Sneddon. 1995.Fourier transforms. Courier Corporation

    1995

  16. [24]

    Weiping Song, Zhiping Xiao, Yifan Wang, Laurent Charlin, Ming Zhang, and Jian Tang. 2019. Session-Based Social Recommendation via Dynamic Graph Attention Networks. InProceedings of the Twelfth ACM International Conference on Web Search and Data Mining. ACM, 555–563

    2019

  17. [25]

    Souza, Diego Mesquita, Samuel Kaski, and Vikas Garg

    Amauri H. Souza, Diego Mesquita, Samuel Kaski, and Vikas Garg. 2022. Provably expressive temporal graph networks. InNeurIPS

    2022

  18. [26]

    Junwei Su, Difan Zou, and Chuan Wu. 2024. PRES: Toward Scalable Memory- Based Dynamic Graph Neural Networks.arXiv preprint arXiv:2402.04284(2024)

    2024 arXiv

  19. [27]

    Yuxing Tian, Yiyan Qi, and Fan Guo. 2023. FreeDyG: Frequency Enhanced Continuous-Time Dynamic Graph Model for Link Prediction. InThe Twelfth International Conference on Learning Representations

    2023

  20. [28]

    Rakshit Trivedi, Hanjun Dai, Yichen Wang, and Le Song. 2017. Know-evolve: Deep temporal reasoning for dynamic knowledge graphs. Ininternational conference on machine learning. PMLR, 3462–3471

    2017

  21. [29]

    Rakshit Trivedi, Mehrdad Farajtabar, Prasenjeet Biswal, and Hongyuan Zha. 2019. DyRep: Learning Representations over Dynamic Graphs. In7th International Conference on Learning Representations. OpenReview.net

    2019

  22. [30]

    Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. 2017. Graph attention networks.arXiv preprint arXiv:1710.10903(2017)

    2017 arXiv

  23. [31]

    Xuhong Wang, Ding Lyu, Mengjian Li, Yang Xia, Qi Yang, Xinwen Wang, Xin- guang Wang, Ping Cui, Yupu Yang, Bowen Sun, et al. 2021. Apan: Asynchronous propagation attention network for real-time temporal graph embedding. InPro- ceedings of the 2021 international conference on ma...

    2021

  24. [32]

    Yanbang Wang, Yen-Yu Chang, Yunyu Liu, Jure Leskovec, and Pan Li. 2021. Inductive Representation Learning in Temporal Networks via Causal Anony- mous Walks. In9th International Conference on Learning Representations. Open- Review.net

    2021

  25. [33]

    Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. 2021. Autoformer: De- composition transformers with auto-correlation for long-term series forecasting. Advances in neural information processing systems34 (2021), 22419–22430

    2021

  26. [34]

    Yuxia Wu, Yuan Fang, and Lizi Liao. 2024. On the Feasibility of Simple Trans- former for Dynamic Graph Modeling. InProceedings of the ACM on Web Confer- ence 2024. 870–880

    2024

  27. [35]

    Yi Xie, Yun Xiong, and Yangyong Zhu. 2020. SAST-GNN: a self-attention based spatio-temporal graph neural network for traffic prediction. InDatabase Systems for Advanced Applications: 25th International Conference, DASFAA 2020, Jeju, South Korea, September 24–27, 2020, Proceedi...

    2020

  28. [36]

    Da Xu, Chuanwei Ruan, Evren Körpeoglu, Sushant Kumar, and Kannan Achan

  29. [37]

    Lu Yi, Jie Peng, Yanping Zheng, Fengran Mo, Zhewei Wei, Yuhang Ye, Yue Zixuan, and Zengfeng Huang. 2025. TGB-Seq Benchmark: Challenging Temporal GNNs with Complex Sequential Dynamics. arXiv:2502.02975 [cs.LG] https://arxiv.org/ abs/2502.02975

    2025 arXiv

  30. [38]

    Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, and Tie-Yan Liu. 2021. Do transformers really perform badly for graph representation?Advances in neural information processing systems34 (2021), 28877–28888

    2021

  31. [39]

    Jiaxuan You, Tianyu Du, and Jure Leskovec. 2022. ROLAND: Graph Learning Framework for Dynamic Graphs. InThe 28th ACM SIGKDD Conference on Knowl- edge Discovery and Data Mining. ACM, 2358–2366

    2022

  32. [40]

    Le Yu, Leilei Sun, Bowen Du, and Weifeng Lv. 2023. Towards better dynamic graph learning: New architecture and unified library.Advances in Neural Information Processing Systems36 (2023), 67686–67700

    2023

  33. [41]

    Mengqi Zhang, Shu Wu, Xueli Yu, Qiang Liu, and Liang Wang. 2022. Dynamic graph neural networks for sequential recommendation.IEEE Transactions on Knowledge and Data Engineering(2022)

    2022

  34. [42]

    Xiaohui Zhang, Yanbo Wang, Xiyuan Wang, and Muhan Zhang. 2024. Efficient Neural Common Neighbor for Temporal Graph Link Prediction.arXiv preprint arXiv:2406.07926(2024)

    2024

  35. [43]

    Ziwei Zhao, Xi Zhu, Tong Xu, Aakas Lizhiyu, Yu Yu, Xueying Li, Zikai Yin, and En- hong Chen. 2023. Time-interval Aware Share Recommendation via Bi-directional Continuous Time Dynamic Graphs. InProceedings of the 46th International ACM SIGIR Conference on Research and Developme...

    2023

  36. [44]

    Yanping Zheng, Lu Yi, and Zhewei Wei. 2024. A survey of dynamic graph neural networks.arXiv preprint arXiv:2404.18211(2024)

    2024 arXiv

  37. [45]

    Tao Zou, Yuhao Mao, Junchen Ye, and Bowen Du. 2024. Repeat-Aware Neighbor Sampling for Dynamic Graph Learning.arXiv preprint arXiv:2405.17473(2024). TIDFormer: Exploiting Temporal and Interactive Dynamics Makes A Great Dynamic Graph Transformer KDD ’25, August 3–7, 2025, Toron...

    2024 arXiv

  38. [2020]

    In8th International Conference on Learning Representations

    Inductive representation learning on temporal graphs. In8th International Conference on Learning Representations. OpenReview.net

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.