Pith. sign in

REVIEW 4 major objections 3 minor 1 cited by

Optimal initialisation and normalisation for GNN fraud detection on the Elliptic Bitcoin dataset is architecture-dependent: GraphSAGE peaks with Xavier alone, GAT with GraphNorm plus Xavier, and GCN is largely unaffected.

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

2026-08-02 20:15 UTC pith:JJPNCEOC

load-bearing objection A careful empirical ablation with a plausible architecture-dependent finding, but the headline ranking rests on small AUPRC gaps, per-architecture (not per-cell) hyperparameters, and no reported error bars. the 4 major comments →

arxiv 2602.23599 v1 pith:JJPNCEOC submitted 2026-02-27 cs.LG

Normalisation and Initialisation Strategies for Graph Neural Networks in Blockchain Anomaly Detection

Dang Sy Duy , Nguyen Duy Chien , Kapil Dev , Jeff Nijsse This is my paper
classification cs.LG
keywords graph neural networksanti-money launderingElliptic Bitcoin datasetGraphNormXavier initialisationclass imbalancetemporal evaluationanomaly detection
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.

This paper asks whether weight initialisation and normalisation choices can be picked once for all graph neural networks used in blockchain anti-money laundering, and answers no. On the Elliptic Bitcoin transaction dataset, the authors compare Xavier initialisation and GraphNorm across GCN, GAT, and GraphSAGE, using temporally ordered splits and AUPRC to handle extreme class imbalance. They find that GraphSAGE performs best with Xavier initialisation alone (AUPRC 0.6678), GAT improves most when GraphNorm is combined with Xavier (+0.055 AUPRC), and GCN is largely insensitive, with its default baseline already best. The practical upshot is that training practices for GNN-based fraud detection should be chosen per architecture rather than assumed to transfer.

Core claim

The central discovery is that the effect of initialisation and normalisation on GNNs for AML is architecture-dependent on Elliptic-like transaction graphs. With temporal splits and repeated sub-sampling, GraphSAGE's peak AUPRC of 0.6678 comes from Xavier initialisation alone; adding GraphNorm slightly reduces it. GAT's best AUPRC rises from 0.6022 to 0.6568 when GraphNorm is combined with Xavier, while Xavier alone lowers its AUC. GCN's baseline configuration already gives its best AUPRC (0.5993), and Xavier or GraphNorm offer no consistent gain. No single strategy dominates all three architectures.

What carries the argument

The argument is carried by a controlled ablation comparing architecture defaults against two training interventions: Xavier initialisation, which preserves activation variance across layers, and GraphNorm, a graph-level normalisation layer applied after message passing that uses graph statistics to stabilise features and reduce over-smoothing. These are applied to three message-passing architectures — GCN, GAT, and GraphSAGE — under a leakage-free temporal split of the Elliptic dataset, with AUPRC (computed over 100 bootstrapped test subsets) as the primary metric. The ablation isolates how each intervention interacts with each architecture's aggregation mechanism.

Load-bearing premise

The comparison assumes that the hyperparameters tuned once per architecture (learning rate, dropout, width, depth) are equally good for every initialisation/normalisation variant; since the headline AUPRC differences are small (at most 0.055), an interaction between hyperparameters and training strategy could reverse the reported ranking.

What would settle it

Re-run the ablation with hyperparameters re-tuned independently for each initialisation/normalisation variant; if GraphSAGE's Xavier-only edge over GraphNorm+Xavier or GAT's GraphNorm advantage disappears or reverses, the architecture-dependence claim is not settled.

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

If this is right

  • For AML pipelines on imbalanced transaction graphs, normalisation and initialisation should be tuned per GNN architecture; a single default recipe will leave performance on the table.
  • GraphSAGE with Xavier initialisation is a strong configuration for illicit-transaction detection, outperforming the prior benchmark on the Elliptic dataset.
  • GAT's attention-based aggregation benefits from graph-level normalisation, suggesting GraphNorm helps stabilise training under skewed degree distributions.
  • GraphNorm can marginally hurt GraphSAGE, so graph-level normalisation is not a free lunch and should be validated before adoption.
  • The proposed evaluation protocol — temporal splits, AUPRC focus, bootstrapped test subsets — provides a fairer basis for comparing GNNs in fraud detection than threshold-dependent F1 at a fixed cutoff.

Where Pith is reading between the lines

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

  • If this architecture-dependence holds more broadly, model-selection pipelines for financial crime should jointly optimise architecture and training hyperparameters rather than fixing one normalisation scheme across models; this is a testable extension.
  • Because the headline gains are small, independently re-tuning hyperparameters for each init/norm variant might shift the rankings; a natural follow-up experiment would check whether the ordering survives.
  • The GAT+GraphNorm synergy may reflect normalisation counteracting heterophily and degree skew; one could test this by evaluating attention-based GNNs on other skewed transaction graphs, such as the Elliptic2 dataset.

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

4 major / 3 minor

Summary. This paper presents an empirical ablation of weight initialisation (Xavier vs. default) and graph normalisation (GraphNorm vs. BatchNorm baseline) for three GNN architectures — GCN, GAT, and GraphSAGE — on the Elliptic Bitcoin transaction dataset. The authors use a temporally ordered 29/10/10 train/validation/test split, train on the full graph, and evaluate primarily with AUPRC. The central claim is that the best training strategy is architecture-dependent: GraphSAGE peaks with Xavier initialisation alone, GAT benefits most from GraphNorm combined with Xavier, and GCN is best left in its baseline configuration. The paper also provides a reproducibility framework and compares with prior published AUPRC results.

Significance. If the architecture-dependent conclusion were established, the paper would offer practical, actionable guidance for AML practitioners and contribute a useful negative result about universal gains from normalisation. The temporal split, AUPRC-based evaluation, and the release of code and logs are genuine strengths. However, the current evidence is not yet sufficient: the reported quantitative differences are small, the claimed repeated-subsampling standard deviations are absent from the main results table, and the hyperparameter protocol does not rule out a confounding interaction between hyperparameters and the init/norm conditions. The central claim therefore needs additional experimental support or a more cautious rewriting.

major comments (4)
  1. [§3.9, Table 2] Section 3.9 states that the authors randomly select 50% of the test nodes 100 times and report 'the mean and standard deviation' for AUPRC and AUC-ROC. However, Table 2 reports only single point values with no standard deviations, confidence intervals, or significance tests. This is load-bearing because the decisive differences are small — e.g., GraphSAGE Xavier-only improves over baseline by 0.0127 AUPRC, and GCN's GraphNorm+Xavier row drops by 0.0551 AUPRC. Without variability estimates, these differences could easily be within sampling noise. The paper should report the full mean ± std (or a compact equivalent) for every cell in Table 2, and ideally perform a paired significance test across the 100 subsamples for the claimed best-vs-second configurations.
  2. [§3.6/Table 1 vs. Table 2] Hyperparameters are optimised once per architecture (Table 1), and the manuscript never states that they were re-tuned for each init/norm cell in Table 2. If the Table 1 values were obtained for the baseline configuration, then the Xavier and GraphNorm+Xavier cells inherit hyperparameters that may be systematically mismatched to those conditions. This is especially relevant because GraphNorm changes activation statistics and therefore interacts with learning rate, dropout, and layer count. The paper's claim to isolate the effect of init/norm presumes an axiom — that one hyperparameter set per architecture is equally appropriate for all init/norm variants — that is not justified. The limitation section (§6.1) acknowledges hardware constraints on tuning but does not address this confound. To support the architecture-dependent conclusion, the authors should either re-tune per cell (at least
  3. [§2.3, RQ2, Table 2] The paper motivates Kaiming/He initialisation in §2.3 and RQ2 asks broadly 'How do initialisation strategies affect optimisation stability and convergence?', yet Table 2 compares only baseline vs. Xavier vs. GraphNorm+Xavier. Kaiming initialisation is never evaluated, despite being explicitly promised as a strategy. Similarly, GraphNorm is never evaluated without Xavier, so the claimed GAT benefit from 'combining GraphNorm with Xavier' cannot be separated into a GraphNorm main effect versus an interaction with Xavier. The title and contributions promise a 'systematic ablation' of initialisation and normalisation strategies; the missing cells should either be added (feasible within the same framework) or the claims and framing should be narrowed accordingly.
  4. [§3.7 / Fig. 1 caption / Table 2] The description of the baseline configuration is internally inconsistent. Section 3.5 says the baseline uses BatchNorm, while Section 3.7 states that 'GraphNorm is applied after each message-passing layer' in each model, and the Fig. 1 caption also lists GraphNorm in the modular design. Table 2's column headings do not specify which normalisation is used in the 'Baseline' and 'Xavier' columns. This ambiguity makes it difficult for a reader to know exactly which configuration is being compared and undermines the reproducibility claim. Please clarify the exact per-cell pipeline and correct the inconsistent narrative.
minor comments (3)
  1. [Abstract / §1] The abstract says 'specifically weight initialisation and normalisation' — remove the stray 'specifically'. The phrase 'such as specifically' is ungrammatical and appears in multiple places.
  2. [Table 2 / §5] The GCN row in Table 2 shows GraphNorm+Xavier at AUPRC 0.5442, a substantial drop from baseline. This is discussed in §4 but not in §5, where the GCN conclusion is only 'limited sensitivity'. Either quantify the drop or temper the 'limited sensitivity' phrasing to acknowledge the clear degradation under GraphNorm.
  3. [Fig. 3 / §3.9] The high-confidence threshold analysis (90th, 99th, 99.9th percentiles) is mentioned in §3.9, but Fig. 3 shows only the 90% threshold. The 99th and 99.9th results are not visible in the figure; consider adding them or explicitly stating they are omitted for space.

Circularity Check

0 steps flagged

No circularity: the paper is an empirical ablation; reported AUPRC values are measured outcomes, not quantities defined via fitted parameters or self-citation chains.

full rationale

The paper's central claim is that optimal initialization/normalization strategy is architecture-dependent, supported by Table 2 AUPRC/AUC numbers for GCN, GAT, and GraphSAGE under baseline, Xavier, and GraphNorm+Xavier configurations. This is an empirical comparison, not a derivation. No fitted parameter is renamed as a prediction: hyperparameters are tuned on a held-out validation set and the test set is strictly held out, as stated in Section 3.6: 'The test set is strictly held out and used only for final evaluation after hyperparameter tuning.' The conclusions are post hoc interpretations of measured metrics, so there is no self-definitional reduction, no fitted-input-called-prediction, and no uniqueness theorem imported from the authors' prior work. The one notable methodological concern—hyperparameters were optimized once per architecture, not per initialization/normalization cell, and Table 2 omits the standard deviations promised in Section 3.9—is a validity/confound issue affecting whether the small AUPRC differences are robust. It is not circularity, because the reported quantities are not constructed to equal the inputs. No load-bearing self-citation appears; references to GraphNorm, Xavier, or prior benchmarks are external published work. The manuscript's own limitations section acknowledges scope constraints but does not expose a circular step. Therefore the circularity score is 0.

Axiom & Free-Parameter Ledger

9 free parameters · 4 axioms · 0 invented entities

The paper introduces no new model components or entities; all assumptions are standard ML/domain assumptions about dataset validity, implementation correctness, and stability of the tuning/evaluation protocol. The most fragile assumption is that per-architecture hyperparameters remain appropriate under all init/norm variants.

free parameters (9)
  • Learning rate (per architecture) = GAT 6.999e-4; GCN 8.475e-4; GraphSAGE 5.302e-4
    Selected by Optuna on validation AUPRC (Table 1); reused across all init/norm conditions, so it is a free parameter of the comparison.
  • Hidden dimension (per architecture) = GAT 148; GCN 211; GraphSAGE 140
    Tuned via Optuna (Table 1).
  • Embedding dimension (per architecture) = GAT 89; GCN 90; GraphSAGE 103
    Tuned via Optuna (Table 1).
  • Dropout rate (per architecture) = GAT 0.2522; GCN 0.2361; GraphSAGE 0.1135
    Tuned via Optuna (Table 1).
  • Number of epochs (per architecture) = GAT 508; GCN 497; GraphSAGE 397
    Tuned via Optuna (Table 1).
  • Number of layers = 2 for all architectures
    Tuned via Optuna (Table 1); shared across configurations.
  • Aggregator (GraphSAGE) = mean
    Tuned via Optuna (Table 1).
  • Weight decay = 5e-4 (fixed)
    Fixed by hand for all models (Section 3.6); affects optimisation and regularisation.
  • Temporal split sizes = 29 train / 10 val / 10 test
    Chosen by hand (Section 3.3); different splits would change comparability with prior benchmarks.
axioms (4)
  • domain assumption The Elliptic dataset labels and 166-dimensional features are reliable, and the 29/10/10 temporal split prevents information leakage.
    All reported AUPRC values depend on this; unknown nodes are retained for message passing but only labeled nodes supervise training (Section 3.3).
  • domain assumption PyTorch Geometric's default reset_parameters and the authors' GraphNorm implementation faithfully realise Xavier/GraphNorm as described.
    The comparison attributes performance differences to init/normalisation, so implementation fidelity is assumed (Sections 3.5, 3.8).
  • ad hoc to paper Hyperparameters tuned once per architecture on validation are also appropriate for alternative init/norm configurations.
    Table 1 gives one hyperparameter set per architecture, reused for Baseline/Xavier/GraphNorm+Xavier; if optimal hyperparameters interact with init/norm, rankings could change.
  • domain assumption Repeated subsampling of 50% of test nodes 100 times yields stable AUPRC estimates.
    Section 3.9 describes this bootstrap, but Table 2 does not report the standard deviations, so stability is assumed.

pith-pipeline@v1.3.0-alltime-deepseek · 9022 in / 11438 out tokens · 100155 ms · 2026-08-02T20:15:44.417732+00:00 · methodology

0 comments
read the original abstract

Graph neural networks (GNNs) offer a principled approach to financial fraud detection by jointly learning from node features and transaction graph topology. However, their effectiveness on real-world anti-money laundering (AML) benchmarks depends critically on training practices such as specifically weight initialisation and normalisation that remain underexplored. We present a systematic ablation of initialisation and normalisation strategies across three GNN architectures (GCN, GAT, and GraphSAGE) on the Elliptic Bitcoin dataset. Our experiments reveal that initialisation and normalisation are architecture-dependent: GraphSAGE achieves the strongest performance with Xavier initialisation alone, GAT benefits most from combining GraphNorm with Xavier initialisation, while GCN shows limited sensitivity to these modifications. These findings offer practical, architecture-specific guidance for deploying GNNs in AML pipelines for datasets with severe class imbalance. We release a reproducible experimental framework with temporal data splits, seeded runs, and full ablation results.

Figures

Figures reproduced from arXiv: 2602.23599 by Dang Sy Duy, Jeff Nijsse, Kapil Dev, Nguyen Duy Chien.

Figure 1
Figure 1. Figure 1: The base model architecture. The evaluated architecture follows a modu [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Convergence behaviour of GCN, GAT, and GraphSAGE on the Elliptic [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Performance comparison of GCN, GAT, and GraphSAGE on the Elliptic [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. From Accuracy to Auditability: A Survey of Determinism in Financial AI Systems

    cs.AI 2026-05 unverdicted novelty 5.0

    Financial AI systems using tabular models, graph networks, and LLM agents exhibit nondeterminism that undermines reproducibility, quantified via experiments on public datasets and addressed by a proposed layered evalu...

Reference graph

Works this paper leans on

26 extracted references · 1 canonical work pages · cited by 1 Pith paper

  1. [1]

    In: Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD)

    Akiba, T., Sano, S., Yanase, T., Ohta, T., Koyama, M.: Optuna: A next- generation hyperparameter optimization framework. In: Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD). pp. 2623–2631 (2019). https://doi.org/10.1145/3292500.3330701

    arXiv 2019

  2. [2]

    In: Proceedings of the 2020 5th International Conference on Machine Learning Technologies

    Alarab, I., Prakoonwit, S., Nacer, M.I.: Competence of graph convolutional net- works for anti-money laundering in bitcoin blockchain. In: Proceedings of the 2020 5th International Conference on Machine Learning Technologies. p. 23–27. ICMLT ’20, Association for Computing Machinery, New York, NY, USA (2020). https://doi.org/10.1145/3409073.3409080

    arXiv 2020

  3. [3]

    arXiv preprint arXiv:1607.06450 (2016), https://arxiv.org/abs/1607.06450 Graph Neural Networks in Blockchain Anomaly Detection 13

    Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016), https://arxiv.org/abs/1607.06450 Graph Neural Networks in Blockchain Anomaly Detection 13

    Pith/arXiv arXiv 2016

  4. [4]

    arXiv preprint arXiv:2404.19109 (2024), https://arxiv.org/abs/2404.19109

    Bellei, C., Xu, M., Phillips, R., Robinson, T., Weber, M., Kaler, T., Leiserson, C.E., Arvind, Chen, J.: The shape of money laundering: Subgraph representation learn- ing on the blockchain with the elliptic2 dataset. arXiv preprint arXiv:2404.19109 (2024), https://arxiv.org/abs/2404.19109

    Pith/arXiv arXiv 2024

  5. [5]

    In: Meila, M., Zhang, T

    Cai, T., Luo, S., Xu, K., He, D., Liu, T.Y., Wang, L.: Graphnorm: A principled approach to accelerating graph neural network training. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 1204–1215. PMLR (2021), https://proceedings.mlr.press/v139/cai21e.html

    2021

  6. [6]

    In: International Conference on Learning Repre- sentations (2018), https://openreview.net/forum?id=rytstxWAW

    Chen, J., Ma, T., Xiao, C.: FastGCN: Fast learning with graph convolutional net- works via importance sampling. In: International Conference on Learning Repre- sentations (2018), https://openreview.net/forum?id=rytstxWAW

    2018

  7. [7]

    arXiv preprint arXiv:2405.19383 (2024), https://arxiv.org/abs/ 2405.19383

    Deprez, B., Vanderschueren, T., Baesens, B., Verdonck, T., Verbeke, W.: Network analytics for anti-money laundering: A systematic literature review and experi- mental evaluation. arXiv preprint arXiv:2405.19383 (2024), https://arxiv.org/abs/ 2405.19383

    arXiv 2024

  8. [8]

    Journal of Machine Learning Research23, 1421– 1478 (2022), https://jmlr.org/papers/v23/22-0567.html

    Dwivedi, V.P., Joshi, C.K., Luu, A.T., Laurent, T., Bengio, Y., Bresson, X.: Bench- marking graph neural networks. Journal of Machine Learning Research23, 1421– 1478 (2022), https://jmlr.org/papers/v23/22-0567.html

    2022

  9. [9]

    arXiv preprint arXiv:1903.02428 (2019), https://arxiv.org/abs/1903.02428

    Fey, M., Lenssen, J.E.: Fast graph representation learning with pytorch geometric. arXiv preprint arXiv:1903.02428 (2019), https://arxiv.org/abs/1903.02428

    Pith/arXiv arXiv 1903

  10. [10]

    In: Teh, Y.W., Titterington, M

    Glorot, X., Bengio, Y.: Understanding the difficulty of training deep feedforward neural networks. In: Teh, Y.W., Titterington, M. (eds.) Proceedings of the Thir- teenth International Conference on Artificial Intelligence and Statistics. Proceed- ings of Machine Learning Research, vol. 9, pp. 249–256. PMLR, Chia Laguna Resort, Sardinia, Italy (2010), http...

    2010

  11. [11]

    Synthesis Lectures on Artificial Intelligence and Machine Learning14, 1–159 (2020)

    Hamilton, W.: Graph representation learning. Synthesis Lectures on Artificial Intelligence and Machine Learning14, 1–159 (2020). https://doi.org/10.2200/ S01045ED1V01Y202009AIM046

    2020

  12. [12]

    arXiv preprint arXiv:1706.02216 (2017), https://arxiv.org/abs/1706.02216

    Hamilton, W.L., Ying, R., Leskovec, J.: Inductive representation learning on large graphs. arXiv preprint arXiv:1706.02216 (2017), https://arxiv.org/abs/1706.02216

    Pith/arXiv arXiv 2017

  13. [13]

    arXiv preprint arXiv:1502.01852 (2015), https://arxiv.org/abs/1502.01852

    He, K., Zhang, X., Ren, S., Sun, J.: Delving deep into rectifiers: Surpassing human- level performance on imagenet classification. arXiv preprint arXiv:1502.01852 (2015), https://arxiv.org/abs/1502.01852

    Pith/arXiv arXiv 2015

  14. [14]

    arXiv preprint arXiv:1502.03167 (2015), https: //arxiv.org/abs/1502.03167

    Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167 (2015), https: //arxiv.org/abs/1502.03167

    Pith/arXiv arXiv 2015

  15. [15]

    arXiv preprint arXiv:1609.02907 (2017), https://arxiv.org/abs/1609

    Kipf, T.N., Welling, M.: Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907 (2017), https://arxiv.org/abs/1609. 02907

    Pith/arXiv arXiv 2017

  16. [16]

    In: 2024 IEEE 21st Consumer Communi- cations & Networking Conference (CCNC)

    Marasi, S., Ferretti, S.: Anti-money laundering in cryptocurrencies through graph neural networks: A comparative study. In: 2024 IEEE 21st Consumer Communi- cations & Networking Conference (CCNC). pp. 272–277 (2024). https://doi.org/ 10.1109/CCNC51664.2024.10454631

    arXiv 2024

  17. [17]

    In: Proceedings of the 26th International Conference on Multimodal Interaction

    Mohamed Selim, A., Bhatti, O.S., Barz, M., Sonntag, D.: Perceived text relevance estimation using scanpaths and gnns. In: Proceedings of the 26th International Conference on Multimodal Interaction. p. 418–427. ICMI ’24, Association for Com- puting Machinery, New York, NY, USA (2024). https://doi.org/10.1145/3678957. 3685736 14 D. Dang et al

    work page doi:10.1145/3678957 2024

  18. [18]

    Scientific Reports14(2024)

    Noval, A., Upadhyay, H., Lagos, L., Soni, J., Prabakar, N.: Spatial-temporal anal- ysis of groundwater well features from neural network prediction of hexavalent chromium concentration. Scientific Reports14(2024). https://doi.org/10.1038/ s41598-024-82297-8

    2024

  19. [19]

    arXiv preprint arXiv:2006.10637 (2020), https://arxiv.org/abs/2006.10637

    Rossi, E., Chamberlain, B., Frasca, F., Eynard, D., Monti, F., Bronstein, M.M.: Temporal graph networks for deep learning on dynamic graphs. arXiv preprint arXiv:2006.10637 (2020), https://arxiv.org/abs/2006.10637

    Pith/arXiv arXiv 2006

  20. [20]

    PLOS ONE 10(3), e0118432 (2015)

    Saito, T., Rehmsmeier, M.: The precision-recall plot is more informative than the ROC plot when evaluating binary classifiers on imbalanced datasets. PLOS ONE 10(3), e0118432 (2015). https://doi.org/10.1371/journal.pone.0118432

    doi:10.1371/journal.pone.0118432 2015

  21. [21]

    arXiv preprint arXiv:1710.10903 (2018), https://arxiv.org/ abs/1710.10903

    Veličković, P., Cucurull, G., Casanova, A., Romero, A., Liò, P., Bengio, Y.: Graph attention networks. arXiv preprint arXiv:1710.10903 (2018), https://arxiv.org/ abs/1710.10903

    Pith/arXiv arXiv 2018

  22. [22]

    In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=XLxhEjKNbXj

    Wang, X., Zhang, M.: GLASS: GNN with labeling tricks for subgraph represen- tation learning. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=XLxhEjKNbXj

    2022

  23. [23]

    arXiv preprint arXiv:1908.02591 (2019), https://arxiv.org/abs/1908.02591

    Weber, M., Domeniconi, G., Chen, J., Weidele, D.K.I., Bellei, C., Robinson, T., Leiserson, C.E.: Anti-money laundering in bitcoin: Experimenting with graph con- volutional networks for financial forensics. arXiv preprint arXiv:1908.02591 (2019), https://arxiv.org/abs/1908.02591

    Pith/arXiv arXiv 1908

  24. [24]

    In: International Conference on Learn- ing Representations (2020), https://openreview.net/forum?id=BJe8pkHFwS

    Zeng,H.,Zhou,H.,Srivastava,A.,Kannan,R.,Prasanna,V.K.:Graphsaint:Graph sampling based inductive learning method. In: International Conference on Learn- ing Representations (2020), https://openreview.net/forum?id=BJe8pkHFwS

    2020

  25. [25]

    In: International Conference on Learning Representations (2020), https://openreview.net/forum? id=rkecl1rtwB

    Zhao, L., Akoglu, L.: Pairnorm: Tackling oversmoothing in gnns. In: International Conference on Learning Representations (2020), https://openreview.net/forum? id=rkecl1rtwB

    2020

  26. [26]

    In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H

    Zhou, K., Huang, X., Li, Y., Zha, D., Chen, R., Hu, X.: Towards deeper graph neural networks with differentiable group normalization. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neu- ral Information Processing Systems. vol. 33, pp. 4917–4928. Curran Asso- ciates, Inc. (2020), https://proceedings.neurips.cc/paper_fi...

    2020