Pith. sign in

REVIEW 4 major objections 5 minor 52 references

Evaluating Loss Functions for Graph Neural Networks: Towards Pretraining and Generalization

T0 review · 4 major / 5 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 The paper claims hybrid losses beat single losses and GIN wins average rank in inductive GNN pretraining, based on a 7-architecture, 30-loss, 21-metric study.

desk verdict The benchmark scale is real, but the headline rankings are an artifact of averaging only top-3 metrics, and the loss the paper praises as Cross-Entropy is actually an MSE denoising objective. read the letter →

arxiv 2506.14114 v1 pith:ZU33BWUT submitted 2025-06-17 cs.LG

Khushnood Abbas , Ruizhe Hou , Zhou Wengang , Dong Shi , Niu Ling , Satyaki Nan , Alireza Abbasi This is my paper
classification cs.LG
keywords graphneuralnetworksunsupervisedrepresentationlearninglossfunctionshybridinductivenodeembeddingsGINpretraining
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 tries to establish an empirical ranking: for unsupervised graph neural network pretraining, hybrid loss functions—weighted combinations of several base objectives—generally outperform single losses, and the best choice of loss depends on the architecture. Across seven GNN families, thirty single and mixed loss functions, three datasets, and twenty-one evaluation metrics, the authors report that GIN is the most consistent architecture in inductive settings, with the combination they call CrossE_L reaching the best average rank (4.95). GAT excels on specific metrics even when its average rank is worse, while MPNN trails the field. The practical stake is that practitioners should treat architecture and loss as a joint design decision rather than picking each separately. The loss called CrossE_L is defined in equation (43) as a mean-squared denoising reconstruction objective, so that label is best read with the equation in hand.

What carries the argument

The load-bearing object is the loss-function family and the rank protocol it is tested under. The five base losses are Contr_L (margin contrastive), PMI_L (cosine similarity aligned to pointwise mutual information), PR_L (PageRank-similarity contrastive), Triplet_L, and CrossE_L, which equation (43) defines as the mean-squared error between a node-feature matrix and a denoised reconstruction of it. Hybrid losses are built as sums of base terms, pairwise products, and triple products with sigmoid-constrained learnable weights, so the comparison includes 30 single-plus-hybrid objectives. The argument is carried by the average-rank tables: every model–loss pair is ranked on each of 21 metrics (node classification, link prediction, adjacency alignment, clustering, and coherence/rank measures), the mean of those ranks is the headline score, and the analysis then filters to top-3 performers per metric.

What would settle it

Rerun the same inductive protocol with a genuine cross-entropy loss in place of the equation-(43) denoising MSE and, separately, remove the all-positive link-prediction recall metric; if GIN + CrossE_L does not remain at the top of the average-rank table, the central ranking fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that optimal unsupervised pretraining is not a property of the architecture alone or the loss alone but of their interaction, and that in inductive settings hybrid objectives are the reliable choice. The authors define five base losses—contrastive, PMI, PageRank-based, triplet, and a reconstruction objective they label CrossE_L—and form hybrids as sigmoid-weighted sums of base terms and their products; they then rank every model–loss pair by average rank over 21 metrics. Their headline result is that GIN + CrossE_L has the lowest average rank (4.95) in inductive transfer, that hybrid losses fill most top positions, that GAT is a specialist with many top-1 wins despite higher average rank, and that MPNN is consistently weak. Equation (43) defines CrossE_L as the mean-squared error between a node-feature matrix and a denoised reconstruction of it, not as a cross-entropy loss, so the narrative label should be read against that definition.

Load-bearing premise

The rankings assume that averaging ranks over the chosen 21 metrics and then keeping only the top three performers per metric is a fair, informative way to compare embeddings.

Editorial extensions

If this is right

  • Practitioners pretraining GNNs inductively should expect multi-objective losses—combinations such as contrastive + PMI or reconstruction + PMI + PageRank—to give more stable embeddings than any single loss.
  • GIN's sum-based aggregator becomes the baseline to beat for unsupervised node representation learning, at least under rank-aggregated evaluation.
  • Task-specific deployments should consult top-1 wins and coverage, not just average rank: GAT with triplet hybrids can be the right choice for a single priority metric.
  • The MPNN framework in its generic form is not sufficient for strong unsupervised pretraining; structural inductive biases are doing the work.
  • Selection decisions should report average rank, coverage, and top-1 wins together, since each alone gives a one-sided view.

Reading between the lines

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

  • Because CrossE_L is actually a denoising MSE objective, the paper's 'cross-entropy' headline is better evidence for reconstruction-based pretraining; a direct comparison with a true cross-entropy loss would separate those readings.
  • The top-3-per-metric filtering favours consistency and may underestimate specialist pairs; recomputing ranks over all combinations, or reweighting metrics, is a natural robustness check that could reorder the leaders.
  • The inductive protocol trains one shared encoder on Cora and CiteSeer and applies it to a 5,000-node Bitcoin subgraph; whether the ranking holds on larger, heterophilous, or feature-rich graphs is an open scaling question.
  • Several of the 21 metrics can be trivially saturated, so removing such metrics is a direct way to test whether the hybrid-loss conclusion stands.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper reports a large-scale empirical comparison of seven GNN encoders and thirty single/hybrid unsupervised loss functions on three datasets, evaluated with twenty-one metrics in both transductive and inductive settings. It claims that hybrid loss functions generally outperform single losses, that GIN with a loss labeled CrossE_L is the most consistent inductive performer, that GAT shows specialized top-1 strengths, and that MPNN typically underperforms. The main evidence is Tables 1 and 2, which are constructed by filtering the top-3 model-loss combinations per metric and then summarizing their average ranks, coverage, and top-1 wins.

Significance. A systematic reference evaluation of the interaction between GNN architectures and unsupervised loss functions would be valuable to practitioners, and the paper's breadth—seven architectures, many loss combinations, three datasets, and two training settings—is an asset. However, the headline rankings are built on an invalid aggregation procedure, the loss called CrossE_L is not cross-entropy, and several of the twenty-one metrics are undefined or degenerate. Because the central claims rest on these flawed summaries, the paper in its current form cannot serve as a reliable reference. Reproducibility is further limited by the absence of code, incomplete hyperparameter reporting, and the undefined 'selfCluster' metric.

major comments (4)
  1. [§5.2.2, Table 2] The Table 2 caption defines AvgRank as the 'mean rank across all metrics', but the reported values are actually averages only over the metrics in which the configuration already appears in the top 3, i.e., over the metrics counted in 'Coverage'. For example, GIN+CrossE_L has Coverage 2 and AvgRank 4.95 = (5.4 + 4.5)/2, exactly the two average-rank entries in Supplementary Tables 31 (LP recall) and 40 (coherence). GCN+Contr_l+PMI_L has Coverage 1 and AvgRank 5.20, matching its single graph-reconstruction-BCE entry in Supplementary Table 36, and SAGE+Triplet_L has Coverage 1 and AvgRank 55.80, matching Supplementary Table 39. This circular top-k averaging makes configurations that appear on very few metrics comparable to configurations that appear on many metrics, and it does not provide any global mean over all twenty-one metrics. The abstract's claim that GIN 'always showed the highest-level average performance' rests directly on this number and is therefore unsupported.
  2. [§5.2.1, Table 1; Supplementary Tables 1 and 10] The transductive headline 'GCN+CrossE_L average rank 1.00 with Coverage 1' is the LP-recall value from Supplementary Table 10, where CrossE_L rows achieve 100% recall on every dataset by predicting all pairs as positive. The full result tables show that the same configurations produce constant majority-class predictions on node classification: Supplementary Table 1 reports every CrossE_L row at 35.24/21.47/71.20 with average ranks around 196–204. Thus the summary's best average rank is driven by a single degenerate metric and directly contradicts the paper's own full results.
  3. [§4.5.3, Eq. (43)] The loss named CrossE_L—and repeatedly described as 'Cross-Entropy loss' in Sections 5.2.1 and 6—is defined in Eq. (43) as the mean squared reconstruction error of a denoising autoencoder, not as a cross-entropy objective. The conclusion that 'GIN with Cross-Entropy loss' is the best inductive configuration is therefore actually about an MSE denoising objective under a misleading name, and the recommendation cannot be transferred to any true cross-entropy loss. The naming should be corrected throughout, or an actual cross-entropy loss should be implemented and evaluated.
  4. [§4.2; Supplementary Tables 20, 40, 41] Several of the 21 metrics are undefined or degenerate, which further undermines the top-3 aggregation. The 'selfCluster' metric is listed in §4.2 without any formal definition, yet it contributes to the 21-metric average and, in Supplementary Tables 20 and 41, most configurations take an identical constant value (about -0.79), so the associated ranks are arbitrary. Coherence is saturated at 100.00 for many GIN rows (Supplementary Table 40), and LP recall rewards the all-positive predictor (Supplementary Tables 10 and 31). Since the summary ranking is computed from which configurations enter the top-3 on such metrics, the reported ordering of model-loss pairs is not a robust conclusion.
minor comments (5)
  1. [§4.5.6] Equations (46)–(48) do not precisely define the hybrid loss family: Eq. (46) is the first-order sum, Eq. (47) is written as a sum of two weight-loss terms rather than a pairwise product, and Eq. (48) contains an ellipsis. The exact set of 30 evaluated combinations and the initialization and optimization of the learnable weights θ should be stated explicitly.
  2. [§4.4] The 'Bitcoin Transaction network' subsection describes the full Elliptic graph, but the experiments use a 5000-node subgraph. The subgraph sampling procedure and the actual graph size should be stated in the main text and reflected in the claims about scale.
  3. [References] Reference pairs [48]/[49] and [51]/[52] are duplicate entries, and several references are cited with incomplete bibliographic data; the reference list should be cleaned and deduplicated.
  4. [§5.2.2, Table 2] The Top1Wins column is not consistent with the component average ranks shown in the supplements: GIN+CrossE_L is credited with 2 top-1 wins even though its two constituent table-average ranks are 5.4 and 4.5, neither of which is rank 1. The counting rule for Top1Wins should be clarified.
  5. [§4.2, Supplementary Tables 12/13 and 33/34] The cosine-adj_corr and dot-adj_corr metrics are reported with identical values in the supplementary tables, suggesting they are redundant in the current implementation. The authors should either distinguish the two metrics or justify including both in the 21-metric average.

Circularity Check

1 steps flagged · score 6.0 of 10

The headline 'AvgRank' in Table 2 is the average of only the top-3-selected metrics, not the mean across all 21 metrics, so the claim that GIN+CrossE_L is the most consistent performer is forced by the aggregation protocol.

  1. fitted input called prediction [Section 5.2.2, Table 2 note and summary; Supplementary Tables 31 and 40]
    "AvgRank: Mean rank across all metrics (lower is better). ... we have filtered only top 3 results based on average rank from inductive results. ... The GIN model with CrossEntropy loss (CrossE_L) achieves the lowest average rank (4.95)"

    The reported AvgRank 4.95 is not the mean across all 21 metrics as the table note defines. In Supplementary Table 31, GIN+CrossE_L has per-metric average rank 5.4 on LP recall; in Table 40, it has 4.5 on coherence. These are the only two metrics where the pair appears in the top-3 (Coverage=2), and (5.4+4.5)/2 = 4.95. The headline 'lowest average rank' is therefore the average of the two ranks already selected by the top-3 filter, i.e., the statistic is constructed from the selection it is supposed to summarize.

full rationale

This is an empirical benchmark paper, not a derivation, so most of the comparison is self-contained: the GNN architectures are defined in Eqs. 31-40, the loss functions in Eqs. 41-48, and the 21 metrics in Eqs. 17-30. The PMI loss does come from the authors' prior work [1], but it is defined and evaluated in this paper alongside independently defined losses, so that self-citation is not load-bearing, and it does not by itself force any conclusion. The one genuinely circular element is the aggregation protocol in Section 5.2.2: the table note defines AvgRank as the mean rank across all metrics, but the actual values are computed only over the metrics in which a configuration already made the top-3. This makes the headline ranking of GIN+CrossE_L an artifact of the selection filter rather than an independent global average. Separately, the loss named CrossE_L is defined in Eq. 43 as a mean-squared denoising reconstruction error, not as cross-entropy; this is a labeling/correctness problem that affects how the abstract's 'Cross-Entropy loss' should be interpreted, but it is not itself a circularity. Because the central empirical claim is undercut by the constructed summary statistic, the score is 6 rather than 0-2.

Assumptions & free parameters 3 free parameters · 4 assumptions · 1 invented entities

The paper's conclusions rely on a set of evaluation design choices: the 21-metric rank aggregation, the representative nature of the small Bitcoin subgraph, the equal-treatment training budget, and the inductive encoder's ability to bridge different input dimensions. The only clearly invented evaluation quantity is the under-specified selfCluster metric. No new physical entities are proposed.

free parameters (3)
  • Learnable hybrid loss weights theta_i, theta_ij, theta_ijk = not reported
    Equations (46) to (48) propose sigmoid-constrained weights for combining base losses; these are optimized during training, so the reported advantage of hybrid losses includes per-task tuning of the combination weights.
  • Margin M in contrastive, triplet, and PageRank losses = not reported
    Sections 4.5.2, 4.5.4, and 4.5.5 define M > 0 but never give its value, despite it controlling the ranking objective.
  • Universal Feature Encoder dimensions (d_h, d_out) = not reported
    Section 4.1 defines the inductive encoder with hidden and output dimensions but the values used in experiments are not stated.
assumptions (4)
  • domain assumption The 21 evaluation metrics, including self-defined and partially formalized metrics (selfCluster, coherence), are valid and equally important measures of embedding quality.
    Every conclusion is a ranking over these metrics; if they are degenerate or redundant, the rankings are not meaningful. Introduced in Section 4.2.
  • domain assumption A 5000-node subgraph of the Bitcoin/Elliptic graph is representative of the full illicit and licit transaction network for transfer learning conclusions.
    Section 4.4 states only a subgraph of 5000 nodes is used; inductive generalization claims rely on this sample.
  • domain assumption Training for 500 epochs with early stopping patience 10 and the same conditions across all datasets and losses gives a fair comparison.
    Section 5 says all experimental conditions were kept the same, but differing loss scales and dataset sizes can make fixed training budgets unfair.
  • domain assumption The Universal Feature Encoder (Eq. 16) can map variable input feature dimensions to a common space so that inductive transfer across Cora, CiteSeer, and Bitcoin is meaningful.
    Section 4.1 assumes that one learned projection from different input dimensions preserves information needed for downstream tasks.
invented entities (1)
  • selfCluster evaluation metric
    purpose: Used as one of 21 metrics to measure internal clusterability, but no formal definition or code is provided.
    Section 4.2 defines selfCluster only by a vague sentence; it appears in Table 20 with constant values near -0.79, and its contribution to rankings is unclear.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Evaluating Loss Functions for Graph Neural Networks: Towards Pretraining and Generalization." pith.science (2026). https://pith.science/paper/ZU33BWUT

@misc{pith2026250614114,
  author       = {Pith},
  title        = {Pith review of: Evaluating Loss Functions for Graph Neural Networks: Towards Pretraining and Generalization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZU33BWUT}},
  note         = {Machine review of arXiv:2506.14114}
}
read the original abstract

Graph Neural Networks (GNNs) became useful for learning on non-Euclidean data. However, their best performance depends on choosing the right model architecture and the training objective, also called the loss function. Researchers have studied these parts separately, but a large-scale evaluation has not looked at how GNN models and many loss functions work together across different tasks. To fix this, we ran a thorough study - it included seven well-known GNN architectures. We also used a large group of 30 single plus mixed loss functions. The study looked at both inductive and transductive settings. Our evaluation spanned three distinct real-world datasets, assessing performance in both inductive and transductive settings using 21 comprehensive evaluation metrics. From these extensive results (detailed in supplementary information 1 \& 2), we meticulously analyzed the top ten model-loss combinations for each metric based on their average rank. Our findings reveal that, especially for the inductive case: 1) Hybrid loss functions generally yield superior and more robust performance compared to single loss functions, indicating the benefit of multi-objective optimization. 2) The GIN architecture always showed the highest-level average performance, especially with Cross-Entropy loss. 3) Although some combinations had overall lower average ranks, models such as GAT, particularly with certain hybrid losses, demonstrated incredible specialized strengths, maximizing the most top-1 results among the individual metrics, emphasizing subtle strengths for particular task demands. 4) On the other hand, the MPNN architecture typically lagged behind the scenarios it was tested against.

Figures

Figures reproduced from arXiv: 2506.14114 by the authors.

Figure 1
Figure 1. This figure shows Average Rank per Model [PITH_FULL_IMAGE:figures/full_fig_p022_1.png] view at source ↗
Figure 2
Figure 2. This figure shows Average Rank per Loss Function. From inductive case top 3 results only. [PITH_FULL_IMAGE:figures/full_fig_p022_2.png] view at source ↗
Figure 3
Figure 3. This is heatmap of model+loss function across metrics in inductive settings. [PITH_FULL_IMAGE:figures/full_fig_p023_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: This figure shows top-1 model combinations. It shows how many times a model and loss function [PITH_FULL_IMAGE:figures/full_fig_p024_4.png]
Figure 5
Figure 5. Figure 5: Model appearance in top 3 rankings across all metrics. [PITH_FULL_IMAGE:figures/full_fig_p024_5.png]
Figure 6
Figure 6. Figure 6: Loss appearance in top 3 rankings across all metrics. [PITH_FULL_IMAGE:figures/full_fig_p025_6.png]
Figure 7
Figure 7. Figure 7: This figure shows Average Rank achieved by models in inductive vs transductive settings. [PITH_FULL_IMAGE:figures/full_fig_p026_7.png]
Figure 8
Figure 8. Figure 8: This figure shows Average Rank achieved by models+loss function in inductive vs transductive settings. [PITH_FULL_IMAGE:figures/full_fig_p026_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 37 canonical work pages

  1. [1]

    Khushnood Abbas, Shi Dong, Alireza Abbasi, and Yong Tang. 2025. Cross-domain inductive applications with unsupervised (dynamic) Graph Neural Networks (GNN): Leveraging Siamese GNN and energy-based PMI optimization. Physica D: Nonlinear Phenomena 476 (2025), 134632. doi:10.1016/j.physd.2025.134632

    work page arXiv 2025

  2. [2]

    Mikhail Belkin and Partha Niyogi. 2003. Laplacian Eigenmaps for Dimensionality Reduction and Data Representation. Neural Comput. 15, 6 (2003), 1373–1396. doi:10.1162/089976603321780317

    doi:10.1162/089976603321780317 2003

  3. [3]

    Shaosheng Cao, Wei Lu, and Qiongkai Xu. 2016. Deep neural networks for learning graph representations. InProceedings of the AAAI conference on artificial intelligence , Vol. 30

    work page 2016

  4. [4]

    Ijeoma Amuche Chikwendu, Xiaoling Zhang, Isaac Osei Agyemang, Isaac Adjei Mensah, Chiagoziem Chima Ukwuoma, and Chukwuebuka Joseph Ejiyi. 2023. A Comprehensive Survey on Deep Graph Representation Learning Methods. J. Artif. Intell. Res. 78 (2023), 287–356. doi:10.1613/JAIR.1.14768

    work page doi:10.1613/jair.1.14768 2023

  5. [5]

    Jun Jin Choong, Xin Liu, and Tsuyoshi Murata. 2020. Optimizing Variational Graph Autoencoder for Community Detection with Dual Optimization. Entropy 22, 2 (2020), 197. doi:10.3390/E22020197

    work page doi:10.3390/e22020197 2020

  6. [6]

    Arash Dehghan-Kooshkghazi, Bogumil Kaminski, Lukasz Krainski, Pawel Pralat, and François Théberge. 2022. Evalu- ating node embeddings of complex networks. J. Complex Networks 10, 4 (2022). doi:10.1093/COMNET/CNAC030

    work page doi:10.1093/comnet/cnac030 2022

  7. [7]

    Iakovos Evdaimon, Giannis Nikolentzos, Michail Chatzianastasis, Hadi Abdine, and Michalis Vazirgiannis. 2024. Neural Graph Generator: Feature-Conditioned Graph Generation using Latent Diffusion Models. CoRR abs/2403.01535 (2024). doi:10.48550/ARXIV.2403.01535

    arXiv doi:10.48550/arxiv.2403.01535 2024

  8. [8]

    Quentin Garrido, Randall Balestriero, Laurent Najman, and Yann LeCun. 2023. RankMe: Assessing the Downstream Performance of Pretrained Self-Supervised Representations by Their Rank. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA (Proceedings of Machine Learning Research, Vol. 202) , Andreas Krause, Emma ...

    work page 2023

Show all 52 references
  1. [9]

    Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. 2017. Neural message passing for quantum chemistry. International Conference on Machine Learning (2017), 1263–1272

    2017

  2. [10]

    Marcos Paulo Silva Gôlo, José Gilberto Barbosa de Medeiros Júnior, Diego Furtado Silva, and Ricardo Marcondes Marcacini. 2024. OLGA: One-cLass Graph Autoencoder. CoRR abs/2406.09131 (2024). doi:10.48550/ARXIV.2406.09131 arXiv:2406.09131

    2024 doi arXiv

  3. [11]

    Martin Grohe. 2020. word2vec, node2vec, graph2vec, X2vec: Towards a Theory of Vector Embeddings of Structured Data. In Proceedings of the 39th ACM SIGMOD-SIGACT-SIGAI Symposium on Principles of Database Systems, PODS 2020, Portland, OR, USA, June 14-19, 2020 , Dan Suciu, Yufei...

    2020

  4. [12]

    Aditya Grover and Jure Leskovec. 2016. node2vec: Scalable Feature Learning for Networks. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (2016)

    2016

  5. [13]

    Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs. Advances in neural information processing systems 30 (2017)

    2017

  6. [14]

    Mario Heidrich, Jeffrey Heidemann, Rüdiger Buchkremer, and Gonzalo Wandosell Fernández de Bobadilla. 2025. ffstruc2vec: Flat, Flexible and Scalable Learning of Node Representations from Structural Identities.CoRR abs/2504.01122 (2025)

    2025 arXiv

  7. [15]

    Binyuan Hui, Pengfei Zhu, and Qinghua Hu. 2020. Collaborative Graph Convolutional Networks: Unsupervised Learning Meets Semi-Supervised Learning. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial...

    2020 doi

  8. [16]

    Haojun Jiang, Jiawei Sun, Jie Li, and Chentao Wu. 2024. LocalGCL: Local-aware Contrastive Learning for Graphs. arXiv preprint arXiv:2402.17345 (2024)

    2024 arXiv

  9. [17]

    Shima Khoshraftar and Aijun An. 2024. A Survey on Graph Representation Learning Methods. ACM Trans. Intell. Syst. Technol. 15, 1 (2024), 19:1–19:55. doi:10.1145/3633518

    2024 doi

  10. [18]

    Dongki Kim, Jinheon Baek, and Sung Ju Hwang. 2022. Graph Self-supervised Learning with Accurate Discrepancy Learning. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, N...

    2022

  11. [19]

    Jinwoo Kim, Olga Zaghen, Ayhan Suleymanzade, Youngmin Ryou, and Seunghoon Hong. 2024. Revisiting Random Walks for Learning on Graphs. ArXiv abs/2407.01214 (2024)

    2024 arXiv

  12. [20]

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

    2016 arXiv

  13. [21]

    Thomas N Kipf and Max Welling. 2016. Variational graph auto-encoders. arXiv preprint arXiv:1611.07308 (2016)

    2016 arXiv

  14. [22]

    Paola Lecca and Michela Lecca. 2023. Graph embedding and geometric deep learning relevance to network biology and structural chemistry. Frontiers Artif. Intell. 6 (2023)

    2023

  15. [23]

    You Li, Bei Lin, Binli Luo, and Ning Gui. 2023. Graph Representation Learning Beyond Node and Homophily. IEEE Trans. Knowl. Data Eng. 35, 5 (2023), 4880–4893. doi:10.1109/TKDE.2022.3146270

    2023

  16. [24]

    Gribova, Vladimir Fedorovich Filaretov, and De-Shuang Huang

    Zhipeng Li, Hailong Su, Xiao bo Zhu, V. Gribova, Vladimir Fedorovich Filaretov, and De-Shuang Huang. 2024. SSPool: A Simple Siamese Framework for Graph Infomax Pooling. IEEE Transactions on Network Science and Engineering 11 (2024), 463–470

    2024

  17. [25]

    Yang Liu, Deyu Bo, Wenxuan Cao, Yuan Fang, Yawen Li, and Chuan Shi. 2025. Graph Positional Autoencoders as Self-supervised Learners. arXiv preprint arXiv:2505.23345 (2025)

    2025 arXiv

  18. [26]

    Yixin Liu, Yizhen Zheng, Daokun Zhang, Vincent C. S. Lee, and Shirui Pan. 2023. Beyond Smoothing: Unsupervised Graph Representation Learning with Edge Heterophily Discriminating. In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023, Thirty-Fifth Conference o...

    2023

  19. [27]

    Weishuo Ma, Yanbo Wang, Xiyuan Wang, and Muhan Zhang. 2024. Reconsidering the Performance of GAE in Link Prediction. CoRR abs/2411.03845 (2024). doi:10.48550/ARXIV.2411.03845 arXiv:2411.03845

    2024 doi arXiv

  20. [28]

    Alexandru Cristian Mara, Jefrey Lijffijt, Stephan Günnemann, and Tijl De Bie. 2022. A Systematic Evaluation of Node Embedding Robustness. In Learning on Graphs Conference, LoG 2022, 9-12 December 2022, Virtual Event (Proceedings of Machine Learning Research, Vol. 198), Bastian...

    2022

  21. [29]

    Ivan Ovinnikov. 2019. Poincar\’e wasserstein autoencoder. arXiv preprint arXiv:1901.01427 (2019)

    2019 arXiv

  22. [30]

    Bryan Perozzi, Rami Al-Rfou, and Steven S. Skiena. 2014. DeepWalk: online learning of social representations. Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining (2014)

    2014

  23. [31]

    Jiezhong Qiu, Yuxiao Dong, Hao Ma, Jian Li, Kuansan Wang, and Jie Tang. 2017. Network Embedding as Matrix Factorization: Unifying DeepWalk, LINE, PTE, and node2vec. Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining (2017)

    2017

  24. [32]

    Guillaume Salha-Galvan and Jiaying Xu. 2025. To Share or Not to Share: Investigating Weight Sharing in Variational Graph Autoencoders. In Companion Proceedings of the ACM on Web Conference 2025, WWW 2025, Sydney, NSW, Australia, 28 April 2025 - 2 May 2025 , Guodong Long, Micha...

    2025

  25. [33]

    Qiaoyu Tan, Ninghao Liu, Xiao Huang, Rui Chen, Soo-Hyun Choi, and Xia Hu. 2022. MGAE: Masked Autoencoders for Self-Supervised Learning on Graphs. CoRR abs/2201.02534 (2022)

    2022 arXiv

  26. [34]

    Jian Tang, Meng Qu, Mingzhe Wang, Ming Zhang, Jun Yan, and Qiaozhu Mei. 2015. LINE: Large-scale Information Network Embedding. Proceedings of the 24th International Conference on World Wide Web (2015)

    2015

  27. [35]

    Anton Tsitsulin, Marina Munkhoeva, and Bryan Perozzi. 2023. Unsupervised Embedding Quality Evaluation. https: //arxiv.org/abs/2305.16562

    2023 arXiv

  28. [36]

    Anton Tsitsulin, Bryan Perozzi, John Palowitch, and Emmanuel Müller. 2023. Graph Clustering with Graph Neural Networks. JMLR 24 (2023). https://arxiv.org/abs/2006.16904

    2023 arXiv

  29. [37]

    Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, Yoshua Bengio, et al. 2017. Graph attention networks. stat 1050, 20 (2017), 10–48550

    2017

  30. [38]

    Hamilton, Pietro Liò, Yoshua Bengio, and R

    Petar Velickovic, William Fedus, William L. Hamilton, Pietro Liò, Yoshua Bengio, and R. Devon Hjelm. 2018. Deep Graph Infomax. CoRR abs/1809.10341 (2018)

    2018 arXiv

  31. [39]

    Zeya Wang and Chenglong Ye. 2024. Deep Clustering Evaluation: How to Validate Internal Clustering Validation Measures. CoRR abs/2403.14830 (2024). doi:10.48550/ARXIV.2403.14830 arXiv:2403.14830

    2024 doi arXiv

  32. [40]

    Weidele, Claudio Bellei, Tom Robinson, and Charles E

    Mark Weber, Giacomo Domeniconi, Jie Chen, Daniel Karl I. Weidele, Claudio Bellei, Tom Robinson, and Charles E. Leiserson. 2019. Anti-Money Laundering in Bitcoin: Experimenting with Graph Convolutional Networks for Financial Forensics. CoRR abs/1908.02591 (2019). http://arxiv.o...

    2019 arXiv

  33. [41]

    Shifeng Xie and Jhony H Giraldo. 2024. Variational Graph Contrastive Learning. arXiv preprint arXiv:2411.07150 (2024)

    2024 arXiv

  34. [42]

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. 2018. How Powerful are Graph Neural Networks? arXiv preprint arXiv:1810.00826 (2018)

    2018 arXiv

  35. [43]

    Hanxuan Yang, Qingchao Kong, and Wenji Mao. 2023. Isomorphic-Consistent Variational Graph Auto-Encoders for Multi-Level Graph Representation Learning. CoRR abs/2312.05519 (2023). doi:10.48550/ARXIV.2312.05519 ACM Comput. Surv., Vol. 37, No. 4, Article 111. Publication date: Au...

    2023 doi arXiv

  36. [44]

    Zhen Yang, Ming Ding, Chang Zhou, Hongxia Yang, Jingren Zhou, and Jie Tang. 2020. Understanding Negative Sampling in Graph Representation Learning. In KDD ’20: The 26th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Virtual Event, CA, USA, August 23-27, 2020 , R...

    2020

  37. [45]

    Jiaxuan You, Rex Ying, and Jure Leskovec. 2019. Position-aware graph neural networks. In International conference on machine learning. PMLR, 7134–7143

    2019

  38. [46]

    Chunhui Zhang and Rui Miao. 2024. Graph contrastive learning with group whitening. InAsian Conference on Machine Learning. PMLR, 1622–1637

    2024

  39. [47]

    Hongyuan Zhang, Pei Li, Rui Zhang, and Xuelong Li. 2022. Embedding graph auto-encoder for graph clustering. IEEE Transactions on Neural Networks and Learning Systems 34, 11 (2022), 9352–9362

    2022

  40. [48]

    Sen Zhang, Qingqing Ye, and Haibo Hu. 2025. Structure-Preference Enabled Graph Embedding Generation under Differential Privacy. ArXiv abs/2501.03451 (2025)

    2025 arXiv

  41. [49]

    Sen Zhang, Qingqing Ye, and Haibo Hu. 2025. Structure-Preference Enabled Graph Embedding Generation under Differential Privacy. CoRR abs/2501.03451 (2025). doi:10.48550/ARXIV.2501.03451 arXiv:2501.03451

    2025 doi arXiv

  42. [50]

    Yichi Zhang and Minh Tang. 2023. A Theoretical Analysis of DeepWalk and Node2vec for Exact Recovery of Community Structures in Stochastic Blockmodels. IEEE Transactions on Pattern Analysis and Machine Intelligence 46 (2023), 1065–1078

    2023

  43. [52]

    Hao Zhu and Piotr Koniusz. 2022. Generalized Laplacian Eigenmaps. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022 , Sanmi Koyejo, S. Moh...

    2022

  44. [53]

    Evaluating Loss Functions for Graph Neural Networks: Towards Pretraining and Generalization

    Yanqiao Zhu, Yichen Xu, Feng Yu, Qiang Liu, Shu Wu, and Liang Wang. 2020. Deep graph contrastive representation learning. arXiv preprint arXiv:2006.04131 (2020). Received 20 June 2025; revised 20 June 2025; accepted 20 June 2025 ACM Comput. Surv., Vol. 37, No. 4, Article 111. ...

    2020 arXiv

Pith tools

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