Pith. sign in

REVIEW 5 major objections 6 minor 50 references

SG-JEPA predicts future node states in latent space with spiking neurons, rivaling supervised accuracy on large dynamic graphs without augmentation, negative sampling, or edge reconstruction.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 15:28 UTC pith:PW26VIAS

load-bearing objection First JEPA-on-dynamic-graphs paper with a coherent spiking encoder; the architecture is worth a look, but the reported experiments rest on an unstated node-absence assumption and some shaky numbers. the 5 major comments →

arxiv 2607.18412 v1 pith:PW26VIAS submitted 2026-07-20 cs.LG cs.NE

Scalable and Efficient Joint Spiking Embedding Predictive Architecture for Large-Scale Dynamic Graphs

Huizhe Zhang , Yuchang Zhu , Huazhen Zhong , Liang Chen , Zibin Zheng This is my paper
classification cs.LG cs.NE
keywords dynamic graph learningself-supervised learningjoint embedding predictive architecturespiking neural networksnode classificationgraph neural networkstemporal graphsscalable representation learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper argues that self-supervised representation learning on large dynamic graphs need not rely on graph augmentation, negative sampling, or edge reconstruction. Instead, SG-JEPA splits each temporal window into context snapshots and a target snapshot, converts each node's history into coarse-to-fine spike-count embeddings using spiking neurons, and trains a predictor to anticipate target nodes' full-precision embeddings directly in latent space. The authors show this predictive objective matches or exceeds supervised dynamic-graph models for node classification on datasets up to 13 million edges, while cutting per-epoch training time by 2.8 to 5.2 times relative to self-supervised baselines. The implication is that latent-space predictive learning, combined with spiking encoders, is a viable route to scalable self-supervised learning on evolving graphs.

Core claim

The central discovery is that joint-embedding predictive learning, previously applied to images and video, can be transplanted to dynamic graphs by partitioning the temporal dimension rather than the input space. Nodes in the final snapshot of a window serve as targets; nodes in earlier snapshots serve as context. A spiking encoder converts each node's history into a coarse-to-fine sequence of spike-count embeddings, and a predictor with learnable pooling produces a representation of the target node. Training uses a contrastive alignment loss with a stop-gradient on the target encoder. On node classification across DBLP, Tmall, and a 13-million-edge Patent graph, the resulting embeddings are

What carries the argument

The engine is the temporal context–target partition: each window of w snapshots designates the last snapshot's nodes as prediction targets and the earlier w−1 snapshots as context. A spiking encoder built from parametric leaky integrate-and-fire neurons converts each node's context history into binary spike trains, then accumulates them into nested spike-count embeddings of increasing temporal resolution. A learnable projection, augmented with positional and sinusoidal time encodings and shared learnable tokens, pools these embeddings through an MLP to predict the target representation. A normalized cross-entropy alignment loss, with stop-gradient on the target encoder, drives training and i

Load-bearing premise

Every target node must yield a sequence of context embeddings from the preceding w−1 snapshots, and the paper does not specify a fallback for nodes absent from earlier snapshots, so the predictive objective is undefined for such nodes and the reported scalability may not transfer to graphs with heavy node churn.

What would settle it

On a dynamic graph with frequent node appearance and disappearance, compute the fraction of target nodes that have complete context histories across the previous w−1 snapshots; if SG-JEPA requires such histories, performance should collapse or training should be undefined when those nodes are included, so replicating the Patent experiment while documenting the absent-node fallback would settle the central scalability claim.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • SG-JEPA reduces training cost substantially: measured per-epoch speedups of 5.2× on DBLP and 2.8× on Tmall over other self-supervised baselines, while avoiding the memory blow-up of edge-level reconstruction that caused an out-of-memory failure for MaskDGNN on the Patent graph.
  • The learned embeddings transfer to node classification without fine-tuning: even at a 40% training ratio, SG-JEPA's Macro-F1 stays within about one point of the best supervised spiking model on all three datasets.
  • Because the spike-count hierarchy is trained to be self-consistent across precisions, the model can serve downstream tasks with fewer context snapshots at inference time, degrading gradually with reduced temporal resolution.
  • The predictive objective converges stably without hand-crafted augmentations, in contrast to reconstruction-based methods whose pretext loss keeps decreasing while downstream performance degrades.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The competitive results suggest the temporal prediction objective itself, not the spiking machinery, may be the main driver; substituting a non-spiking context encoder (e.g., average pooling) would isolate the contribution of spikes, and the authors' ablation with IF/LIF neurons indicates learnable membrane dynamics account for about 1.5–2.0 F1 points.
  • The paper's claim of avoiding negative sampling sits uneasily with its InfoNCE-style loss, whose denominator contrasts each target against other nodes in the window; 'no negative sampling' is better read as 'no explicit, engineered negative-mining or augmentation,' not as a contrastive-free objective.
  • If temporal partitioning requires every target node to have embeddings across the prior w−1 snapshots, then the method's scalability to genuinely evolving graphs depends on an unspecified fallback for absent nodes; a practitioner should ask how missing-node contexts are handled before adoption.
  • The coarse-to-fine spike-count design hints at a broader application: the same nested hierarchy could enable multi-granularity inference on resource-limited devices, trading a few F1 points for large reductions in spike count—an avenue the paper only sketches.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper introduces SG-JEPA, a self-supervised joint-embedding predictive architecture for discrete-time dynamic graphs. The model partitions snapshots into temporal windows; target nodes are those in the last snapshot of a window, and context nodes are the same nodes in preceding snapshots. A GNN target encoder produces full-precision embeddings, a PLIF spiking context encoder converts the historical embeddings into cumulative spike-count embeddings, and an MLP predictor with an InfoNCE loss and stop-gradient aligns predicted and target representations. Experiments on DBLP, Tmall, and Patent report Macro-/Micro-F1 against full-precision, self-supervised, and spiking baselines, together with training-efficiency measurements, sensitivity analysis, and ablations. The paper claims competitive or superior performance relative to supervised spiking baselines and clear gains over self-supervised baselines, while avoiding negative sampling, graph augmentation, and edge reconstruction.

Significance. If the claims are substantiated, SG-JEPA would be a meaningful step: it brings the JEPA predictive principle to dynamic graphs, combines it with spiking neurons for coarse-to-fine temporal encoding, and presents evidence that a non-contrastive, non-generative objective can scale to a 13M-edge graph without labels. The architecture is clearly specified and the experimental design is broad, including multiple baselines, three datasets, efficiency comparisons, sensitivity analysis, and ablation studies. The central idea is timely and worth pursuing. However, the current manuscript does not ship code or data, omits key implementation details, and contains several internal inconsistencies that prevent full verification of the central claims; the claimed scalability and performance therefore remain conditional.

major comments (5)
  1. [§4.1, §4.3, Eq. (10)] The temporal partitioning assumes that every target node v at snapshot iw has a well-defined context sequence z_{v,1},...,z_{v,w-1}. Since the paper defines V_t as the set of nodes active at time t, nodes that are absent from earlier context snapshots have no z_{v,t}. The manuscript never states how such nodes are handled. On Tmall and Patent, node churn is common, so this omission affects the training set definition, the loss, and the reported scaling results. The authors must specify a fallback (e.g., masking, zero-padding, or dropping such nodes) and discuss the resulting bias. This is load-bearing for the method's applicability and reproducibility.
  2. [§4.3, Eq. (11)] Equation (11) defines c_{v,t} = sum_{t=1}^{w-1} s_{v,t}, so the right-hand side is independent of t and all c_{v,t} would be identical. This contradicts the intended coarse-to-fine hierarchy and the later use of prefix subsequences c_{v,1:t}. The equation should be a prefix sum, e.g., c_{v,t} = sum_{tau=1}^{t} s_{v,tau}. Without this correction, the nested spike-count embedding mechanism is not well defined.
  3. [§4.5, Eq. (15)] The InfoNCE loss sums over all target nodes u in V(i) in the denominator. For the reported large-scale graphs, this is O(|V(i)|^2) per loss term and incompatible with the claimed memory scalability unless negative sampling or in-batch negatives are used. The text says SG-JEPA avoids negative sampling, but InfoNCE is still a contrastive objective that treats other nodes as negatives. The paper needs to specify whether V(i) is subsampled, whether in-batch negatives are used, and how the loss is computed efficiently on the Patent dataset. This directly affects the scalability and efficiency claims.
  4. [Tables 2 and 3] Several entries appear to be typos and undermine the baseline comparisons. In Table 2, SpikeNet on Patent at 60% training ratio reports Macro-F1 94.03, whereas the 40% and 80% ratios report 83.92 and 94.20, respectively; all other methods on Patent are around 84. In Table 3, TGAT on Tmall at 60% reports Micro-F1 93.32, whereas the 40% entry is 62.05 and the corresponding Macro-F1 is 57.61. Also, the CLDG entry '6.63±0.40.7' is malformed. These values must be corrected and re-verified; if they are genuine, an explanation is needed.
  5. [§5.3, §5 (experimental setup)] The paper does not provide code, data, or a complete set of default hyperparameters. Section 5.3 refers to default settings for window size, number of layers, number of sampled neighbors, and threshold, but these defaults are never listed. Training details such as learning rate, batch size, number of epochs, and hardware are also absent. This makes the efficiency and scaling results non-reproducible and prevents readers from assessing the sensitivity analysis meaningfully. Please add a full experimental configuration and release code or at minimum a detailed configuration table.
minor comments (6)
  1. [§4.4, Eq. (13)] The temporal encoding of the target node is written as e_w, but the target is at snapshot iw. Use e_{iw} or define the notation for a general window.
  2. [§4.6] The inference step states that the final context sequence is {c_{v,(k-1)w+2}, ..., c_{v,kw}}. This indexing appears inconsistent with the window definition in §4.1, where the context snapshots are (k-1)w+1 through kw-1. Please clarify the intended indices.
  3. [§5.1] The claim that the average performance gap between SG-JEPA and SiGNN is 'only 0.085%' is unclear. Table 2 shows differences above 0.5 Macro-F1 in several settings. Specify whether this is a relative gap, a particular training ratio, or an average over a subset.
  4. [§4.6] The sentence 'spiking count embeddings and full-precision target embeddings are jointly optimized to predict each other' is imprecise: the target embedding is predicted from the context, and the stop-gradient operation prevents the target encoder from being directly trained by the prediction loss. Rephrase to avoid implying symmetric optimization.
  5. [Table 4] The label 'w/ostopgrad' is a typo; it should read 'w/o stopgrad'. Also, the Δ column reports only a single aggregate number; specify how the per-metric drops are averaged.
  6. [General] There are minor grammatical issues, e.g., 'It incurs substantial computational overhead' in the abstract should be 'This incurs...', and 'Conventional' is capitalized mid-sentence in Section 6.

Circularity Check

0 steps flagged

No significant circularity; SG-JEPA's objective is standard self-distillation and results are benchmarked externally. Minor self-citations are contextual, not load-bearing.

full rationale

The central derivation—temporal partitioning (§4.1), spiking context encoding (§4.3), predictor (§4.4), and InfoNCE objective with stop-gradient (§4.5)—does not fit any parameter to a target label and then rename the fit as a prediction. The target representation z_{v,iw} is produced by a GNN encoder from the final snapshot of each window and is used only as a self-supervised regression target with stop-gradient; the same encoder's earlier outputs feed the context SNN. This is the standard JEPA/BYOL-style self-distillation loop, not a circular derivation of the empirical results. Performance is measured against external baselines (JODIE, EvolveGCN, TGAT, ROLAND, CLDG, MaskDGNN, GC-SNN, SpikeNet, Dy-SIGN, Delay-DSGN, SiGNN) on public datasets, so the headline comparisons are not forced by the paper's own equations. Self-citations [25, 26, 44, 46] appear only as contextual examples in related work and do not carry the derivation. Two non-circular weaknesses are noted: (i) Eq. (11) writes c_{v,t} = sum_{t=1}^{w-1} s_{v,t}, which is independent of t, so the claimed 'coarse-to-fine' hierarchy is not actually defined by the equation as written; and (ii) §4.1–4.3 assume z_{v,1..w-1} exist for every target node, but no fallback is specified for nodes absent from earlier snapshots, which matters on Tmall/Patent. These are correctness/reproducibility concerns, not circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

No new physical or theoretical entities are introduced; all components (PLIF, GraphSAGE, InfoNCE, random-walk PE, sinusoidal TE) come from prior work. The main unsupported load is the unstated node-absence handling and the choice of several hyperparameters whose defaults are not reported.

free parameters (5)
  • Temporal window size w = not reported (swept in Fig. 5)
    Defines the context-target partition and the number of spike-count levels; performance depends on it (Fig. 5).
  • Fixed-size neighbor sampling count = not reported (swept in Fig. 5)
    Neighbor sampling controls the subgraph used for both target and context encoding; only a sensitivity range is shown, no default.
  • Number of GNN layers L = not reported (swept in Fig. 5)
    Depth of the target encoder; sensitivity analysis shows non-monotonic performance with depth.
  • InfoNCE temperature tau = not reported
    Sharpness of the discriminative loss in Eq. (15); standard hyperparameter but values are not stated.
  • Spiking neuron threshold V_th and PLIF membrane parameters = not reported
    Threshold controls spike sparsity and is shown in Fig. 5 to affect performance; exact settings are not given.
axioms (5)
  • domain assumption PLIF neuron dynamics (Eqs. 1-3) and surrogate gradient (Eq. 4) are valid for temporal graph representation learning.
    The paper adopts PLIF and surrogate-gradient training from prior work; the efficiency and performance claims depend on spiking temporal coding being useful here.
  • ad hoc to paper Every target node at snapshot iw has a well-defined sequence of embeddings at snapshots (i-1)w+1 through iw-1.
    Eq. (10) sums over t=1..w-1 for every target node; no fallback is given for nodes absent in earlier snapshots, which is common in dynamic graphs.
  • domain assumption InfoNCE with in-batch negatives is a valid surrogate for predictive alignment and prevents representation collapse.
    This is an empirical choice from [18]; the paper assumes discriminative alignment transfers to downstream node classification and relies on stop-gradient for stability.
  • domain assumption Fixed-size neighbor sampling preserves the temporal-structural evolution needed for prediction.
    The paper samples a fixed number of neighbors per node and treats the resulting subgraphs as sufficient context for predicting future node representations.
  • domain assumption Matryoshka-style nested projection (Eq. 12) is a valid way to extract coarse-to-fine predictions from spike-count embeddings.
    The design borrows from Matryoshka representation learning; the paper provides no formal argument that this nesting improves prediction.

pith-pipeline@v1.3.0-alltime-deepseek · 15170 in / 13260 out tokens · 107208 ms · 2026-08-01T15:28:16.126747+00:00 · methodology

0 comments
read the original abstract

Dynamic graph learning aims to capture evolving structural and semantic patterns in real-world systems, such as fraud detection and recommender systems. Due to the scarcity of labeled data in real-world dynamic graphs, recent studies have introduced generative or contrastive paradigms (e.g., masked graph autoencoders or graph contrastive learning) to generate task-agnostic graph embeddings. However, these methods typically rely on complex edge-level reconstruction objectives and tailored graph augmentation strategies. This incurs substantial computational overhead when scaling to large-scale dynamic graphs. In this paper, we propose SG-JEPA, a joint spiking embedding predictive architecture for large-scale dynamic graphs. In contrast to existing self-supervised methods, SG-JEPA partitions nodes into context and target sets along the temporal dimension to learn embeddings that are predictive of each other via additional spatial-temporal information. Furthermore, through encoding sequential inputs into coarse-to-fine spike count embeddings, spiking neurons enable SG-JEPA to adapt to the varying computational constraints of downstream tasks. Extensive experiments demonstrate that SG-JEPA achieves competitive or even superior performance over discriminative baselines on node classification, while effectively scaling to the dynamic graph with 13 million edges. SG-JEPA avoids the complex machinery (negative sampling, graph augmentations, edge-level reconstruction, etc.), resulting in superior training efficiency and memory scalability compared with prior self-supervised dynamic graph baselines.

Figures

Figures reproduced from arXiv: 2607.18412 by Huazhen Zhong, Huizhe Zhang, Liang Chen, Yuchang Zhu, Zibin Zheng.

Figure 1
Figure 1. Figure 1: Illustration of self-supervised learning frameworks for DGNNs. (a) Contrastive-based methods construct multiple [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The workflow of the SG-JEPA. resolutions, while maintaining the computational efficiency and sparsity benefits of SNNs. 4.4 Target Representation Prediction Given the spike count embeddings from the context encoder, the target representation predictor takes the context representations with multiple precision levels to align the full-precision target repre￾sentation. Inspired by Matryoshka representation le… view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of per-epoch training time and Micro [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Parameter sensitivity analysis on the DBLP, Tmall and Patent datasets. 5.3 Parameter Sensitivity Analysis We evaluate the robustness of SG-JEPA with respect to four key hyperparameters: window size, the number of GNN layers, the num￾ber of sampled neighbors and the membrane potential threshold of the spiking neuron. For each hyperparameter, we vary one factor while keeping the others fixed at their default… view at source ↗
Figure 4
Figure 4. Figure 4: Learning curve of three self-supervised methods. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

50 extracted references · 11 linked inside Pith

  1. [1]

    Mohamed Abdelfattah and Alexandre Alahi. 2024. S-jepa: A joint embedding predictive architecture for skeletal action recognition. InEuropean Conference on Computer Vision. Springer, 367–384

    2024

  2. [2]

    Mohammad Ali Alomrani, Mahdi Biparva, Yingxue Zhang, and Mark Coates

  3. [3]

    Uri Alon and Eran Yahav. 2020. On the bottleneck of graph neural networks and its practical implications.arXiv preprint arXiv:2006.05205(2020)

    Pith/arXiv arXiv 2020

  4. [4]

    Mido Assran, Adrien Bardes, David Fan, Quentin Garrido, Russell Howes, Matthew Muckley, Ammar Rizvi, Claire Roberts, Koustuv Sinha, Artem Zho- lus, et al. 2025. V-jepa 2: Self-supervised video models enable understanding, prediction and planning.arXiv preprint arXiv:2506.09985(2025)

    Pith/arXiv arXiv 2025

  5. [5]

    Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. 2023. Self-supervised learning from images with a joint-embedding predictive architecture. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 15619–15629

    2023

  6. [6]

    Adrien Bardes, Quentin Garrido, Jean Ponce, Xinlei Chen, Michael Rabbat, Yann LeCun, Mido Assran, and Nicolas Ballas. 2023. V-jepa: Latent video prediction for visual representation learning.arXiv preprint arXiv:2404.08471(2023)

    Pith/arXiv arXiv 2023

  7. [7]

    Delong Chen, Mustafa Shukor, Theo Moutakanni, Willy Chung, Jade Yu, Te- jaswi Kasarla, Allen Bolourchi, Yann LeCun, and Pascale Fung. 2025. Vl-jepa: Joint embedding predictive architecture for vision-language.arXiv preprint arXiv:2512.10942(2025)

    arXiv 2025

  8. [8]

    Dong Chen, Shuai Zheng, Muhao Xu, Zhenfeng Zhu, and Yao Zhao. 2025. SiGNN: A spike-induced graph neural network for dynamic graph representation learning. Pattern Recognition158 (2025), 111026

    2025

  9. [9]

    Ke-Jia Chen, Linsong Liu, Linpu Jiang, and Jingqiang Chen. 2023. Self-supervised dynamic graph representation learning via temporal subgraph contrast.ACM Transactions on Knowledge Discovery from Data18, 1 (2023), 1–20

    2023

  10. [10]

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. InInterna- tional conference on machine learning. PmLR, 1597–1607

    2020

  11. [11]

    Vijay Prakash Dwivedi, Anh Tuan Luu, Thomas Laurent, Yoshua Bengio, and Xavier Bresson. 2021. Graph neural networks with learnable structural and positional representations.arXiv preprint arXiv:2110.07875(2021)

    Pith/arXiv arXiv 2021

  12. [12]

    Jason K Eshraghian, Max Ward, Emre O Neftci, Xinxin Wang, Gregor Lenz, Girish Dwivedi, Mohammed Bennamoun, Doo Seok Jeong, and Wei D Lu. 2023. Training spiking neural networks using lessons from deep learning.Proc. IEEE111, 9 (2023), 1016–1054

    2023

  13. [13]

    Wei Fang, Zhaofei Yu, Yanqi Chen, Timothée Masquelier, Tiejun Huang, and Yonghong Tian. 2021. Incorporating learnable membrane time constant to en- hance learning of spiking neural networks. InProceedings of the IEEE/CVF inter- national conference on computer vision. 2661–2671

    2021

  14. [14]

    Zhengcong Fei, Mingyuan Fan, and Junshi Huang. 2023. A-jepa: Joint-embedding predictive architecture can listen.arXiv preprint arXiv:2311.15830(2023)

    Pith/arXiv arXiv 2023

  15. [15]

    ZhengZhao Feng, Rui Wang, TianXing Wang, Mingli Song, Sai Wu, and Shuibing He. 2025. A comprehensive survey of dynamic graph neural networks: Models, frameworks, benchmarks, experiments and challenges.IEEE Transactions on Knowledge and Data Engineering(2025)

    2025

  16. [16]

    Data File. 2001. Lessons, Insights and Methodological Tools, ".NBER Working Paper8498 (2001), 40

    2001

  17. [17]

    Mengzhou Gao, Xinxun Zhang, Pengfei Jiao, Tianpeng Li, and Zhidong Zhao. 2025. DVGMAE: Self-Supervised Dynamic Variational Graph Masked Autoencoder. IEEE Transactions on Neural Networks and Learning Systems(2025)

    2025

  18. [18]

    Michael Gutmann and Aapo Hyvärinen. 2010. Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. InProceedings of the thirteenth international conference on artificial intelligence and statistics. JMLR Workshop and Conference Proceedings, 297–304

    2010

  19. [19]

    Ehsan Hajiramezanali, Arman Hasanzadeh, Krishna Narayanan, Nick Duffield, Mingyuan Zhou, and Xiaoning Qian. 2019. Variational graph recurrent neural networks.Advances in neural information processing systems32 (2019)

    2019

  20. [20]

    Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs.Advances in neural information processing systems30 (2017)

    2017

  21. [21]

    Yiming He, Xiang Li, Zhongying Zhao, Haobing Liu, Peilan He, and Yanwei Yu. 2025. MaskDGNN: Self-Supervised Dynamic Graph Neural Networks with Activeness-aware Temporal Masking. InProceedings of the Thirty-Fourth Interna- tional Joint Conference on Artificial Intelligence. 2892–2900

    2025

  22. [22]

    Linpu Jiang, Ke-Jia Chen, and Jingqiang Chen. 2021. Self-supervised dynamic graph representation learning via temporal subgraph contrast.arXiv preprint arXiv:2112.08733(2021)

    Pith/arXiv arXiv 2021

  23. [23]

    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. 1269–1278

    2019

  24. [24]

    Aditya Kusupati, Gantavya Bhatt, Aniket Rege, Matthew Wallingford, Aditya Sinha, Vivek Ramanujan, William Howard-Snyder, Kaifeng Chen, Sham Kakade, Prateek Jain, et al. 2022. Matryoshka representation learning.Advances in Neural Information Processing Systems35 (2022), 30233–30249

    2022

  25. [25]

    Jintang Li, Ruofan Wu, Xinzhou Jin, Boqun Ma, Liang Chen, and Zibin Zheng

  26. [26]

    Jintang Li, Zhouxin Yu, Zulun Zhu, Liang Chen, Qi Yu, Zibin Zheng, Sheng Tian, Ruofan Wu, and Changhua Meng. 2023. Scaling up dynamic graph representation learning via spiking neural networks. InProceedings of the AAAI conference on artificial intelligence, Vol. 37. 8588–8596

    2023

  27. [27]

    Xianhang Li, Chen Huang, Chun-Liang Li, Eran Malach, Josh Susskind, Vimal Thilak, and Etai Littwin. 2025. Rethinking JEPA: Compute-Efficient Video SSL with Frozen Teachers.arXiv preprint arXiv:2509.24317(2025)

    arXiv 2025

  28. [28]

    Yuanfu Lu, Xiao Wang, Chuan Shi, Philip S Yu, and Yanfang Ye. 2019. Temporal network embedding with micro-and macro-dynamics. InProceedings of the 28th ACM international conference on information and knowledge management. 469– 478

    2019

  29. [29]

    Yao Ma, Ziyi Guo, Zhaocun Ren, Jiliang Tang, and Dawei Yin. 2020. Stream- ing graph neural networks. InProceedings of the 43rd international ACM SIGIR conference on research and development in information retrieval. 719–728

    2020

  30. [30]

    Aldo Pareja, Giacomo Domeniconi, Jie Chen, Tengfei Ma, Toyotaro Suzumura, Hiroki Kanezashi, Tim Kaler, Tao Schardl, and Charles Leiserson. 2020. Evolvegcn: Evolving graph convolutional networks for dynamic graphs. InProceedings of the AAAI conference on artificial intelligence, Vol. 34. 5363–5370

    2020

  31. [31]

    Joakim Skarding, Bogdan Gabrys, and Katarzyna Musial. 2021. Foundations and modeling of dynamic networks using dynamic graph neural networks: A survey. iEEE Access9 (2021), 79143–79168

    2021

  32. [32]

    Li Sun, Zhenhao Huang, Qiqi Wan, Hao Peng, and Philip S Yu. 2024. Spiking graph neural network on riemannian manifolds.Advances in Neural Information Processing Systems37 (2024), 34025–34055

    2024

  33. [33]

    Sheng Tian, Ruofan Wu, Leilei Shi, Liang Zhu, and Tao Xiong. 2021. Self- supervised representation learning on dynamic graphs. InProceedings of the 30th ACM international conference on information & knowledge management. 1814–1823

    2021

  34. [34]

    Yanbang Wang, Yen-Yu Chang, Yunyu Liu, Jure Leskovec, and Pan Li. 2021. Inductive representation learning in temporal networks via causal anonymous walks.arXiv preprint arXiv:2101.05974(2021)

    Pith/arXiv arXiv 2021

  35. [35]

    Zhiqiang Wang, Jianghao Wen, and Jianqing Liang. 2025. Delay-DSGN: A Dy- namic Spiking Graph Neural Network with Delay Mechanisms for Evolving Graph. InForty-second International Conference on Machine Learning. Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Huizhe Zhang, Yuchang Zhu, Huazhen Zhong, Liang Chen, and Zibin Zheng

    2025

  36. [36]

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

  37. [37]

    Mingkun Xu, Yujie Wu, Lei Deng, Faqiang Liu, Guoqi Li, and Jing Pei. 2021. Exploiting spiking dynamics with spatial-temporal feature normalization in graph learning.arXiv preprint arXiv:2107.06865(2021)

    Pith/arXiv arXiv 2021

  38. [38]

    Sihan Xu, Ziqiao Ma, Wenhao Chai, Xuweiyi Chen, Weiyang Jin, Joyce Chai, Saining Xie, and Stella X Yu. 2025. Next-Embedding Prediction Makes Strong Vision Learners.arXiv preprint arXiv:2512.16922(2025)

    arXiv 2025

  39. [39]

    Yiming Xu, Bin Shi, Teng Ma, Bo Dong, Haoyi Zhou, and Qinghua Zheng. 2023. Cldg: Contrastive learning on dynamic graphs. In2023 IEEE 39th International Conference on Data Engineering (ICDE). IEEE, 696–707

    2023

  40. [40]

    Man Yao, Guangshe Zhao, Hengyu Zhang, Yifan Hu, Lei Deng, Yonghong Tian, Bo Xu, and Guoqi Li. 2023. Attention spiking neural networks.IEEE transactions on pattern analysis and machine intelligence45, 8 (2023), 9393–9410

    2023

  41. [41]

    Nan Yin, Mengzhu Wang, Zhenghan Chen, Giulia De Masi, Huan Xiong, and Bin Gu. 2024. Dynamic spiking graph neural networks. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 16495–16503

    2024

  42. [42]

    Jiaxuan You, Tianyu Du, and Jure Leskovec. 2022. ROLAND: graph learning framework for dynamic graphs. InProceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining. 2358–2366

    2022

  43. [43]

    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

  44. [44]

    Huizhe Zhang, Jintang Li, Yuchang Zhu, Liang Chen, and Zibin Zheng. 2025. GT- SVQ: A Linear-Time Graph Transformer for Node Classification Using Spiking Vector Quantization.arXiv preprint arXiv:2504.11840(2025)

    arXiv 2025

  45. [45]

    Yifan Zhu, Fangpeng Cong, Dan Zhang, Wenwen Gong, Qika Lin, Wenzheng Feng, Yuxiao Dong, and Jie Tang. 2023. Wingnn: Dynamic graph neural networks with random gradient aggregation window. InProceedings of the 29th ACM SIGKDD conference on knowledge discovery and data mining. 3650–3662

    2023

  46. [46]

    Zulun Zhu, Jiaying Peng, Jintang Li, Liang Chen, Qi Yu, and Siqiang Luo. 2022. Spiking graph convolutional networks.arXiv preprint arXiv:2205.02767(2022)

    Pith/arXiv arXiv 2022

  47. [47]

    Zulun Zhu, Kai Wang, Haoyu Liu, Jintang Li, and Siqiang Luo. 2024. Topology- monitorable contrastive learning on dynamic graphs. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 4700–4711. Received 20 February 2007; revised 12 March 2009; accepted 5 June 2009

    2024

  48. [2020]

    Inductive representation learning on temporal graphs.arXiv preprint arXiv:2002.07962(2020)

    Pith/arXiv arXiv 2002

  49. [2022]

    DyG2Vec: Efficient representation learning for dynamic graphs.arXiv preprint arXiv:2210.16906(2022)

    Pith/arXiv arXiv 2022

  50. [2024]

    State space models on temporal graphs: A first-principles study.Advances in Neural Information Processing Systems37 (2024), 127030–127058

    2024