Pith. sign in

REVIEW 3 major objections 5 minor 51 references

NodeJEPA: Structure-Conditioned Latent Prediction for Node-Level Graph Self-Supervised Learning

T0 review · 3 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read This paper claims that node-level graph self-supervised learning can be done by predicting the latent representations of masked local neighborhoods, and that two such JEPA-style variants tie for the best average rank across five benchmarks.

desk verdict A solid, reproducible node-level JEPA recipe with an honest matched protocol, but the paper's own '4/5 datasets' claim contradicts Table 1, and the fixed baseline hyperparameters make the average-rank advantage look fragile. read the letter →

arxiv 2608.04381 v1 pith:XBXLNLQB submitted 2026-08-05 cs.LG cs.AI

Tinghe Zhang , Jian Xu , Jiaheng Chen , Jiaxing Li , Yucheng Xiao , Qiang Wang This is my paper · ORCID
classification cs.LGcs.AI
keywords graphself-supervisedlearningjoint-embeddingpredictivearchitecturenode-levelrepresentationlatentpredictionmaskedego-subgraphpartitionmaskingvariance-covarianceregularizationfew-shotnodeclassification
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

This paper sets out to bring joint-embedding predictive architectures (JEPA) to node-level graph representation learning: instead of contrasting augmented views or reconstructing masked attributes in input space, the model hides structure-aware k-hop ego-subgraphs and predicts, in latent space, the representations a teacher encoder would produce for the hidden nodes. Across five node-classification benchmarks, the resulting method (NodeJEPA) and a cached-partition variant (PatchJEPA) tie for the best average rank among six self-supervised methods and keep their advantage under few-shot probes. The paper argues that latent prediction over neighborhoods is a strong alternative to contrastive and generative graph SSL, and that the granularity of the masking unit is the main accuracy-versus-efficiency lever.

What carries the argument

The carrying mechanism is the node-level JEPA objective: a context encoder must predict the latent embedding of masked nodes or patches from visible context, using stop-gradient targets produced by an EMA-updated teacher, so the learning signal lives in representation space rather than input space. The two variants named in the paper, NodeJEPA (k-hop ego-subgraph masking with a structure-conditioned predictor) and PatchJEPA (one-time METIS partition with pooled patch embeddings), share this objective and differ only in the masking unit, which lets the paper attribute accuracy and efficiency differences to masking granularity. Collapse prevention comes from a variance-covariance regularizer plus a sketched isotropic-Gaussian penalty, and an optional curriculum raises mask ratio and hop radius during training.

What would settle it

Re-run the five-benchmark comparison with per-dataset hyperparameter tuning for DGI, GraphMAE, BGRL, and CCA-SSG (or with a different shared encoder such as a graph attention network), and check whether NodeJEPA and PatchJEPA still tie for the best average self-supervised rank and still finish first or second on four of five datasets; if not, the central claim would not survive.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that JEPA-style latent prediction transfers from images and whole graphs to individual nodes once the masking unit and predictor conditioning are chosen appropriately. NodeJEPA masks k-hop ego-subgraphs with a learned mask token, trains a GCN context encoder against stop-gradient targets from an EMA-updated target encoder, and conditions the predictor on cheap structural descriptors such as PageRank, degree, clustering coefficient, and spectral coordinates. With VICReg-style variance-covariance and sketched isotropic-Gaussian regularization to block collapse, it attains the best average self-supervised rank (2.4) on five benchmarks, tied with PatchJEPA, which replaces the per-step k-hop search with a cached METIS partition. The authors report that both variants finish first or second on four of five datasets, that the gains persist with 5 to 20 labels per class, and that ablations make context isolation, cosine prediction, and variance-covariance regularization the load-bearing default ingredients.

Load-bearing premise

The reported edge depends on the baselines being evaluated under exactly the same encoder, splits, and hyperparameter choices the authors selected; if the baselines were tuned separately or a different encoder were used, the average-rank lead could shrink or disappear.

Editorial extensions

If this is right

  • If the claim holds, node-level graph SSL no longer needs hand-crafted augmentations, negative pairs, or input reconstruction: predicting masked neighborhoods in latent space suffices for strong downstream probes.
  • Masking granularity becomes the primary design axis, with NodeJEPA preferred on medium graphs and PatchJEPA, whose METIS partition is cached once, the scalable choice on large hub-heavy graphs such as ogbn-arxiv.
  • The few-shot results imply the learned representations remain useful under extreme label scarcity, the setting that most motivates unsupervised pretraining.
  • The ablation findings suggest follow-up recipes should keep target isolation and cosine prediction, and should treat structural descriptors as optional rather than required.

Reading between the lines

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

  • Since removing the explicit structural descriptors did not hurt accuracy, the paper implicitly leaves open that an even simpler predictor, relying only on restricted message passing, might capture nearly all the structural signal; testing that would clarify how much conditioning data the predictor genuinely needs.
  • The VICReg ablation is dataset-dependent (removing it drops Amazon-Photo accuracy but raises Coauthor-Physics accuracy), which suggests per-dataset regularization strength, not a fixed recipe, could push the reported averages higher.
  • The evaluation is confined to homophilous graphs, so the natural next probe is whether latent neighborhood prediction transfers to heterophilous or low-label networks, where the masked-neighborhood pretext may behave differently.
  • The efficiency gap on ogbn-arxiv invites a testable extension: degree-capped or cached k-hop expansion should close most of the gap while keeping fine-grained masking, a change the paper notes but does not implement.
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 / 5 minor

Summary. The paper proposes NodeJEPA and PatchJEPA, two joint-embedding predictive architectures for node-level graph self-supervised learning. NodeJEPA masks structure-aware k-hop ego-subgraphs and predicts stop-gradient target embeddings from a context encoder, conditioning the predictor on structural descriptors and adding VICReg-style and SIGReg-style regularizers. PatchJEPA replaces the k-hop masking unit with a cached METIS patch partition. The authors evaluate both variants under a matched linear-probing protocol on five node classification benchmarks, reporting that NodeJEPA and PatchJEPA tie for the best average self-supervised rank (2.4 of 6) and that the advantage persists in few-shot probes. They also report ablations, representation-collapse diagnostics, and wall-clock efficiency measurements on ogbn-arxiv.

Significance. If the empirical claims are robust, the paper makes a useful contribution by demonstrating that JEPA-style latent prediction can be applied at the node level and by identifying masking granularity as a concrete accuracy-efficiency axis. The manuscript has notable strengths: the matched protocol with shared encoder family and splits, paired significance tests on shared seeds, a public code release, explicit ablations, and measured wall-clock costs. These are meaningful assets. However, the headline claim rests on average ranks computed among only six self-supervised methods, and the baseline configurations used in the matched protocol appear to be unusually weak for some methods. Because the central claim is empirical and rank-based, the calibration of the baselines is load-bearing and must be addressed before the conclusions can be accepted.

major comments (3)
  1. [Section 4.1, Tables 1, 5, 6] The matched protocol fixes one hyperparameter set per method, but the reported baseline numbers are far below typical operating points for these algorithms on the same Shchur splits. For example, Table 1 reports GraphMAE at 70.76±2.92 on Amazon-Computers and 86.05±0.52 on Amazon-Photo, and DGI at 56.96±18.09 and 63.62±13.00; the DGI standard deviations indicate unstable rather than representative runs. Since the headline result is average rank among only six self-supervised methods, a few points of improvement in two or three baselines would change the rank ordering and can erase the reported 2.4 average. No sensitivity analysis, per-dataset baseline tuning, or comparison against published baseline numbers is supplied. The claim that accuracy differences reflect the pretraining objective rather than baseline configuration is therefore not yet supported. The authors should tune baselines per dataset with validation-based selection, report published best numbers alongside their matched runs, or demonstrate that the average-rank advantage survives plausible baseline improvements.
  2. [Figure 1 and Section 1 vs. Table 1] The text and Figure 1 state that NodeJEPA and PatchJEPA each finish first or second on four of five datasets. Table 1 does not support this for PatchJEPA: on ogbn-arxiv its accuracy is 68.86, behind BGRL (68.87) and CCA-SSG (68.85), giving rank 3, and on Coauthor-Physics it ranks 6 among the self-supervised methods. Only NodeJEPA finishes first or second on four datasets; PatchJEPA does so on three. The average-rank value 2.4 is arithmetically correct, but the per-variant claim and the caption of Figure 1 must be corrected.
  3. [Section 5, Table 2] The ablation labeled D shows that removing the VICReg-style variance-covariance term improves Coauthor-Physics accuracy from 90.57 to 92.67, while the introduction and Section 5 describe variance-covariance regularization as a contributing ingredient of the recommended recipe. The later remark that the preferred strength is dataset-dependent is appropriate, but the default configuration is not justified against this result; please state explicitly how the default was chosen, for example by average validation rank rather than per-dataset optimum, and discuss the Coauthor-Physics result as a genuine tension rather than a side note.
minor comments (5)
  1. [Abstract and Section 1] There is a formatting typo: 'NodeJEP A' and 'we proposeNodeJEP A' should be 'NodeJEPA' with correct spacing.
  2. [Table 1 and Table 7] The CCA-SSG row renders as '88.56±0.0793.06±0.1568.85±0.19'; the entries for Coauthor-CS, Coauthor-Physics, and ogbn-arxiv should be separated by spaces or columns.
  3. [Section 1 contributions] The text says 'five benchmarks and five baselines,' but Table 1 lists four self-supervised baselines plus a supervised reference; the count should be clarified to avoid conflating the supervised GCN with the SSL baselines.
  4. [Section 7] The sentence 'Neither extension changes the core claim' is an assertion without supporting evidence; it would be more accurate to say the reported experiments do not test those extensions and that their effect is unknown.
  5. [Section 4.2 and Figure 7] The claim that both variants are 'statistically match the label-supervised GCN reference' on Amazon-Computers is based on a paired test with p=0.956 for NodeJEPA, but the corresponding PatchJEPA comparison has p=0.199; please state the p-values or at least qualify the claim consistently.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the claimed results are empirical comparisons against external baselines and public benchmarks, with no load-bearing step that reduces to a fit or to a self-citation.

full rationale

The paper's central claims are empirical: NodeJEPA and PatchJEPA are trained with a latent-prediction objective and evaluated by linear-probe accuracy and few-shot transfer on five public benchmark graphs (Section 4, Table 1). The average-rank headline is a measured outcome of a matched protocol, not a construction that defines the methods in terms of the baselines. No parameter is fitted to the headline metric and then renamed as a prediction: the method-specific hyperparameters in Table 6 are fixed across datasets, and the ablations in Section 5 vary design choices rather than retrofitting targets. PatchJEPA is explicitly disclosed as a comparison variant that shares the encoder, EMA target network, and regularization recipe with NodeJEPA and changes only the masking unit, so its comparison isolates the stated design axis rather than importing a result by definition. Citations to prior JEPA work (LeCun 2022; Assran et al. 2023; Skenderi et al. 2023; Balestriero and LeCun 2025) are external to this paper's authors and do not carry the paper's own empirical claims; there are no load-bearing self-citations by the present authors. The concern that fixed baseline hyperparameters may understate DGI, BGRL, or GraphMAE is a question of experimental fairness and generalization, not circular reasoning: the paper's derivation chain does not reduce to its inputs or to any self-referential definition. Accordingly no circular step is identified.

Assumptions & free parameters 7 free parameters · 3 assumptions · 0 invented entities

The central claim is empirical, so the ledger captures hand-picked hyperparameters and evaluation assumptions rather than mathematical axioms. No new physical or mathematical entities are introduced; the only new artifacts are the two architecture variants.

free parameters (7)
  • NodeJEPA mask ratio schedule = 0.2 to 0.5 over first 50 epochs
    Chosen by hand to control pretraining difficulty; affects the pretraining objective's strength.
  • NodeJEPA hop radius schedule = 1 to 2 hops over first 50 epochs
    Chosen by hand; determines the size and structure of each masked ego-subgraph.
  • EMA momentum = 0.996 to 0.999 (NodeJEPA), 0.996 (PatchJEPA)
    Standard for JEPA-style methods; controls target encoder update speed.
  • Regularization weights = lambda_var=0.2, lambda_cov=0.2, lambda_sig=0.02
    Chosen by hand; balance collapse prevention against predictive loss.
  • SIGReg projection count = 256 slices
    Chosen by hand for the sketched isotropic-Gaussian penalty.
  • METIS patch count = 64/64/64/128/256 per dataset
    Chosen by hand; sets PatchJEPA's masking granularity.
  • Patch context/target count = 1 context, 4 target patches
    Chosen by hand; follows the Graph-JEPA-style recipe.
assumptions (3)
  • domain assumption The five evaluated datasets (all homophilous: co-purchase, co-authorship, citation) are representative of the node classification settings where graph SSL is used.
    The paper's limitations section (Section 7) acknowledges weaker-homophily settings are untested; the central claim of competitiveness is inferred from these datasets.
  • domain assumption Linear probing of frozen embeddings is a faithful measure of representation quality.
    The evaluation protocol (Section 4.1) uses linear probing to isolate representation quality, but this assumes linear separability is the right downstream proxy.
  • domain assumption The matched protocol is fair to all baselines: a 3-layer GCN with 256 hidden units and the chosen baseline hyperparameters are competitive for each method.
    The conclusion that differences in accuracy are attributable to the pretraining objective depends on baseline hyperparameters not being under-tuned; the weakest assumption section elaborates.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NodeJEPA: Structure-Conditioned Latent Prediction for Node-Level Graph Self-Supervised Learning." pith.science (2026). https://pith.science/paper/XBXLNLQB

@misc{pith2026260804381,
  author       = {Pith},
  title        = {Pith review of: NodeJEPA: Structure-Conditioned Latent Prediction for Node-Level Graph Self-Supervised Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XBXLNLQB}},
  note         = {Machine review of arXiv:2608.04381}
}
read the original abstract

Self-supervised learning on graphs is largely shaped by contrastive methods that depend on carefully designed augmentations, and by generative methods that reconstruct node attributes in the input space. Both paradigms can entangle representations with low-level input statistics rather than with relational structure. Joint-embedding predictive architectures (JEPA) instead learn by predicting latent targets rather than reconstructing inputs. Recent work has explored this idea for graph-level representation learning, but how to design JEPA-style objectives for node-level tasks, and which structural signals the predictor should condition on, remains less clear. We present NodeJEPA, a joint-embedding predictive architecture for node-level graph self-supervised learning. NodeJEPA masks structure-aware k-hop ego-subgraphs and trains a context encoder to predict the latent representations of the masked nodes. These targets come from an EMA-updated target encoder with stop-gradient. A structure-conditioned predictor integrates spectral and centrality descriptors through cross-attention. Variance, covariance, and Laplacian spectral regularizers help stabilize the embedding geometry, and an optional curriculum gradually increases masking difficulty during training. Because prediction occurs in latent space, NodeJEPA does not rely on input reconstruction or hand-crafted graph augmentations. We evaluate NodeJEPA on standard node classification benchmarks under linear probing and fine-tuning protocols, and conduct ablations on masking, prediction, and regularization design choices. Our study offers a practical recipe for node-level JEPA-style latent prediction on graphs, and clarifies when structural conditioning helps representation learning. Code, configurations, and evaluation scripts are publicly available at https://github.com/OliverZ-dot/Node-Jepa.

Figures

Figures reproduced from arXiv: 2608.04381 by the authors.

Figure 1
Figure 1. Average rank among six self-supervised meth [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of NodeJEPA (left) and PatchJEPA (right). Both use an EMA teacher–student encoder pair and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. NodeJEPA ablation results. Removing struc [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figures from the paper (7 more)
Figure 5
Figure 5. Figure 5: Per-epoch wall-clock time on ogbn-arxiv (me [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Linear-probe accuracy across all five benchmarks (same data as Table [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Left: average rank (lower better). Middle: SSL baselines beaten at [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Few-shot accuracy vs. labels per class on all five datasets. [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: NodeJEPA ablation results on Amazon-Photo and Coauthor-Physics. [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: Top: effective rank of frozen embeddings. Bottom: mean per-dimension standard deviation. NodeJEPA [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]
Figure 11
Figure 11. Figure 11: Top: accuracy vs. wall-clock time. Middle: total training time on four medium graphs. Bottom: per-epoch [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 36 canonical work pages

  1. [1]

    Open Review , volume=

    A Path Towards Autonomous Machine Intelligence Version 0.9.2, 2022-06-27 , author=. Open Review , volume=

    work page 2022

  2. [2]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    work page

  3. [3]

    2309.16014 , archivePrefix=

    Graph-level Representation Learning with Joint-Embedding Predictive Architectures , author=. 2309.16014 , archivePrefix=

    arXiv

  4. [4]

    International Conference on Learning Representations , year=

    Deep Graph Infomax , author=. International Conference on Learning Representations , year=

  5. [5]

    Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

    GraphMAE: Self-Supervised Masked Graph Autoencoders , author=. Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

    work page

  6. [6]

    International Conference on Learning Representations , year=

    Large-Scale Representation Learning on Graphs via Bootstrapping , author=. International Conference on Learning Representations , year=

    work page

  7. [7]

    Advances in Neural Information Processing Systems , volume=

    From Canonical Correlation Analysis to Self-supervised Graph Neural Networks , author=. Advances in Neural Information Processing Systems , volume=

    work page

  8. [8]

    International Conference on Learning Representations , year=

    Semi-Supervised Classification with Graph Convolutional Networks , author=. International Conference on Learning Representations , year=

Show all 51 references
  1. [9]

    SIAM Journal on Scientific Computing , volume=

    A Fast and High Quality Multilevel Scheme for Partitioning Irregular Graphs , author=. SIAM Journal on Scientific Computing , volume=

  2. [10]

    International Conference on Learning Representations , year=

    VICReg: Variance-Invariance-Covariance Regularization for Self-Supervised Learning , author=. International Conference on Learning Representations , year=

  3. [11]

    International Conference on Learning Representations , year=

    Graph Attention Networks , author=. International Conference on Learning Representations , year=

  4. [12]

    Advances in Neural Information Processing Systems , volume=

    Open Graph Benchmark: Datasets for Machine Learning on Graphs , author=. Advances in Neural Information Processing Systems , volume=

  5. [13]

    1811.05868 , archivePrefix=

    Pitfalls of Graph Neural Network Evaluation , author=. 1811.05868 , archivePrefix=

    arXiv

  6. [14]

    Advances in Neural Information Processing Systems , volume=

    Bootstrap Your Own Latent: A New Approach to Self-Supervised Learning , author=. Advances in Neural Information Processing Systems , volume=

  7. [15]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Masked Autoencoders Are Scalable Vision Learners , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  8. [16]

    Advances in Neural Information Processing Systems , volume=

    Graph Contrastive Learning with Augmentations , author=. Advances in Neural Information Processing Systems , volume=

  9. [17]

    Advances in Neural Information Processing Systems , volume=

    Inductive Representation Learning on Large Graphs , author=. Advances in Neural Information Processing Systems , volume=

  10. [18]

    International Conference on Machine Learning , pages=

    A Simple Framework for Contrastive Learning of Visual Representations , author=. International Conference on Machine Learning , pages=

  11. [19]

    Proceedings of the Web Conference 2021 , pages=

    Graph Contrastive Learning with Adaptive Augmentation , author=. Proceedings of the Web Conference 2021 , pages=

    2021

  12. [20]

    Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

    What's Behind the Mask: Understanding Masked Graph Modeling for Graph Autoencoders , author=. Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

  13. [21]

    2511.08544 , archivePrefix=

    LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics , author=. 2511.08544 , archivePrefix=

    arXiv

  14. [22]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Exploring Simple Siamese Representation Learning , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  15. [23]

    International Conference on Machine Learning , pages=

    Barlow Twins: Self-Supervised Learning via Redundancy Reduction , author=. International Conference on Machine Learning , pages=

  16. [24]

    arXiv preprint arXiv:1807.03748 , year=

    Representation Learning with Contrastive Predictive Coding , author=. arXiv preprint arXiv:1807.03748 , year=

    arXiv

  17. [25]

    Advances in Neural Information Processing Systems , volume=

    Mean Teachers are Better Role Models: Weight-Averaged Consistency Targets Improve Semi-Supervised Deep Learning Results , author=. Advances in Neural Information Processing Systems , volume=

  18. [26]

    Proceedings of the 26th International Conference on Machine Learning , pages=

    Curriculum Learning , author=. Proceedings of the 26th International Conference on Machine Learning , pages=

  19. [27]

    International Conference on Learning Representations , year=

    How Powerful are Graph Neural Networks? , author=. International Conference on Learning Representations , year=

  20. [28]

    Stanford InfoLab Technical Report , year=

    The PageRank Citation Ranking: Bringing Order to the Web , author=. Stanford InfoLab Technical Report , year=

  21. [29]

    Nature , volume=

    Collective Dynamics of `Small-World' Networks , author=. Nature , volume=

  22. [30]

    International Conference on Learning Representations , year=

    Understanding Dimensional Collapse in Contrastive Self-Supervised Learning , author=. International Conference on Learning Representations , year=

  23. [31]

    International Conference on Machine Learning , pages=

    Graph Contrastive Learning Automated , author=. International Conference on Machine Learning , pages=

  24. [32]

    IEEE Transactions on Neural Networks and Learning Systems , volume=

    A Comprehensive Survey on Graph Neural Networks , author=. IEEE Transactions on Neural Networks and Learning Systems , volume=

  25. [33]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    Self-Supervised Learning of Graph Neural Networks: A Unified Review , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

  26. [34]

    Proceedings of the 20th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

    DeepWalk: Online Learning of Social Representations , author=. Proceedings of the 20th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

  27. [35]

    International Conference on Learning Representations , year=

    Strategies for Pre-training Graph Neural Networks , author=. International Conference on Learning Representations , year=

  28. [36]

    2006.04131 , archivePrefix=

    Deep Graph Contrastive Representation Learning , author=. 2006.04131 , archivePrefix=

    2006 arXiv

  29. [37]

    International Conference on Machine Learning , pages=

    Contrastive Multi-View Representation Learning on Graphs , author=. International Conference on Machine Learning , pages=

  30. [38]

    Proceedings of the 26th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

    GCC: Graph Contrastive Coding for Graph Neural Network Pre-Training , author=. Proceedings of the 26th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

  31. [39]

    International Conference on Learning Representations , year=

    InfoGraph: Unsupervised and Semi-supervised Graph-Level Representation Learning via Mutual Information Maximization , author=. International Conference on Learning Representations , year=

  32. [40]

    IEEE International Conference on Data Mining , pages=

    Sub-graph Contrast for Scalable Self-Supervised Graph Representation Learning , author=. IEEE International Conference on Data Mining , pages=

  33. [41]

    Hou, Zhenyu and He, Yufei and Cen, Yukuo and Liu, Xiao and Dong, Yuxiao and Kharlamov, Evgeny and Tang, Jie , booktitle=

  34. [42]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Momentum Contrast for Unsupervised Visual Representation Learning , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  35. [43]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Emerging Properties in Self-Supervised Vision Transformers , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  36. [44]

    Chiang, Wei-Lin and Liu, Xuanqing and Si, Si and Li, Yang and Bengio, Samy and Hsieh, Cho-Jui , booktitle=

  37. [45]

    Neural Computation , volume=

    Laplacian Eigenmaps for Dimensionality Reduction and Data Representation , author=. Neural Computation , volume=

  38. [46]

    International Conference on Machine Learning , pages=

    Neural Message Passing for Quantum Chemistry , author=. International Conference on Machine Learning , pages=

  39. [47]

    Advances in Neural Information Processing Systems , volume=

    Beyond Homophily in Graph Neural Networks: Current Limitations and Effective Designs , author=. Advances in Neural Information Processing Systems , volume=

  40. [48]

    Pei, Hongbin and Wei, Bingzhe and Chang, Kevin Chen-Chuan and Lei, Yu and Yang, Bo , booktitle=

  41. [49]

    Grover, Aditya and Leskovec, Jure , booktitle=

  42. [50]

    NeurIPS Workshop on Bayesian Deep Learning , year=

    Variational Graph Auto-Encoders , author=. NeurIPS Workshop on Bayesian Deep Learning , year=

  43. [51]

    Geometric Deep Learning: Going beyond

    Bronstein, Michael M and Bruna, Joan and LeCun, Yann and Szlam, Arthur and Vandergheynst, Pierre , journal=. Geometric Deep Learning: Going beyond

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.