Pith. sign in

REVIEW 4 major objections 4 minor 55 references

LMAE4Eth: Generalizable and Robust Ethereum Fraud Detection by Exploring Transaction Semantics and Masked Graph Embedding

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Encoding Ethereum transactions as sentences lifts fraud detection

desk verdict Serious empirical method paper with an overstated abstract; the random-split leakage concern is real but partly controlled, and the isolated-split result suggests the core finding survives for the baselines tested. read the letter →

arxiv 2509.03939 v1 pith:SN5Z7Y3L submitted 2025-09-04 cs.CR cs.LG

Yifan Jia , Yanbin Wang , Jianguo Sun , Ye Tian , Peng Qian This is my paper
classification cs.CRcs.LG
keywords Ethereumfrauddetectiontransactionsemanticsmaskedgraphautoencodercontrastivelearninglanguagemodelpretrainingphishingcross-attentionfusionlayer-neighborsampling
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 claims that Ethereum fraud detection improves when raw numeric transaction records are first converted into short attribute-value sentences (for example, "amount: 0.02, direction: 0, time: 15942...") and then read by a language model pre-trained with a masked-token plus contrastive objective. The same accounts are simultaneously embedded by a masked graph autoencoder that reconstructs node features rather than graph structure, using expert-crafted account statistics as initial features, and the two views are merged by a cross-attention fusion network. Across three public Ethereum datasets the combined method beats 21 baselines, with F1-score gains of 10.24%, 8.85%, and 6.49% over the strongest competitor on MulDiGraph, B4E, and SPN. If the claim holds, it means transaction text can supply semantic cues that numerical sequences miss, and node-level fraud detection is better served by feature-reconstructing autoencoders than by structure-reconstructing ones, especially when labeled phishing accounts are scarce.

What carries the argument

Three components carry the argument. (1) TxCLM, the transaction-token contrastive language model: a BERT-style Transformer that converts each transaction tuple into a sentence of attribute-value tokens and is pre-trained with masked transaction modeling plus a token-aware contrastive objective, making token embeddings more discriminative and reducing anisotropy from homogeneous records. (2) MAGAE, the masked account graph autoencoder: it applies layer-neighbor sampling (LABOR) to partition the transaction graph, masks a random subset of node features, and reconstructs them with scaled cosine error — reconstructing node features rather than graph structure is what makes it node-task oriented

What would settle it

Retrain the language branch on the identical transaction tuples serialized as plain numbers without the text labels ('amount:', 'direction:', 'timestamp:'), keeping masking, contrastive loss, graph branch, and fusion unchanged. If F1 on B4E does not drop materially below the reported 0.8143, the claimed semantic advantage is not coming from textualization.

Watch

Extended reading notes

Core claim

The paper's central claim is that Ethereum account activity has two complementary views — transaction semantics visible only when numerical records are written as language, and interaction structure visible in the account graph — and that fusing both detects fraud better than either alone. LMAE4Eth textualizes each transaction as attribute-value tokens (amount, direction, timestamp, addresses excluded), trains a Transformer with masked-language modeling plus a token-aware contrastive loss, and runs a masked graph autoencoder that reconstructs masked node features (initialized with 22 expert features) using scaled cosine error under layer-neighbor sampling. A cross-attention fusion network me

Load-bearing premise

The load-bearing premise is that rendering transaction tuples as attribute-value text such as "amount: 4.3, direction: 1" gives a language model real semantic signal beyond the same numbers fed in raw form.

Editorial extensions

If this is right

  • Fraud patterns in Ethereum can be surfaced through language-model pretraining on textualized transaction histories, so even accounts with very few labels get usable representations.
  • Masked node-feature reconstruction is more aligned with node-level fraud detection than structure reconstruction; graph methods that only reconstruct edges or topology will underperform on account classification.
  • Layer-neighbor sampling reduces the sampled edge budget by about an order of magnitude on a 3-million-node graph, making the graph view scalable without sacrificing F1.
  • Fusing semantic, graph, and expert views through cross-attention outperforms simple addition or linear combination by 5–7% F1, so the views are complementary rather than redundant.
  • Under isolated connected-component splits — where training and test accounts share no graph connectivity — the method still leads by roughly 5–10% F1, implying it generalizes beyond local neighborhoods.

Reading between the lines

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

  • If the linguistic representation is the main driver, the same textualization recipe should transfer to other blockchain ecosystems (e.g., Bitcoin or Tron) and to fine-grained transaction intent classification, not just binary fraud labels.
  • The token-aware contrastive objective plausibly generalizes to other highly homogeneous financial sequences; a cheap test is to apply it to credit-card or bank-transfer logs with the same attribute-value serialization.
  • The graph view adds most for accounts that cannot be judged from transaction text alone (short histories, few tokens). A testable extension: measure LMAE4Eth's margin over the language-only variant as a function of account transaction count; if the margin grows for short histories, the complementarity story is confirmed.
  • A streaming or incremental variant could update semantic and graph views as new transactions arrive, since the two branches are pre-trained separately; the paper lists batch-mode operation as its own limitation.
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 / 4 minor

Summary. The paper proposes LMAE4Eth, a self-supervised framework for Ethereum fraud account detection that combines a transaction-token contrastive language model (TxCLM), a masked account graph autoencoder with layer-neighbor sampling (LABOR-MAGAE), and a cross-attention fusion network (CAFN). Transaction records are converted into attribute-value text, and a BERT-family model is pretrained with masked language modeling and a token-aware contrastive objective. Concurrently, a graph autoencoder reconstructs masked node features initialized with expert-engineered statistics, using LABOR sampling for scalability. The two views are fused with learnable cross-attention tokens and fine-tuned for account classification. The authors report F1 improvements of 10.24%, 8.85%, and 6.49% over the strongest baselines on MulDiGraph, B4E, and SPN, respectively, with additional ablations, architecture-replacement studies, hyperparameter sensitivity analyses, and an isolated connected-component robustness check.

Significance. If the empirical claims hold, LMAE4Eth would be a strong new state of the art for Ethereum fraud detection, with a novel and plausible mechanism: combining linguistic representations of transactions with masked node-feature reconstruction and cross-attention fusion. The paper's strengths include a large baseline suite (21 methods), three datasets, a code link, component-level ablations, an isolated-split control experiment, and efficiency measurements for the sampling strategy. The central 'consistently outperforms all baselines' claim, however, is not yet established under a leakage-free protocol because the main comparison uses random node splits and the isolated-split control covers only 7 of 21 baselines under a different class-balance regime. The abstract also overstates the measured improvements. The overall direction is publishable after substantial revision of the evaluation protocol and reporting.

major comments (4)
  1. [Section VI-A, Table III and Section VI-D, Table VIII] The main SOTA claim rests on random node-based splits (7:1:2) of the transaction graph. In a transaction graph this creates edges between training and test accounts, so supervised GNNs and the fine-tuned fusion module can benefit from transductive label/structural leakage. The isolated connected-component split in Table VIII is the right control, but it covers only 7 of the 21 baselines and also changes the task by sampling a 2:1 benign-to-phishing ratio rather than the original class distribution. Therefore, the statement that LMAE4Eth 'consistently outperforms all baseline methods across three Ethereum datasets' is not yet established under a leakage-free protocol. Please run the full baseline suite under isolated splits, or explicitly justify why the excluded baselines cannot exploit random-split leakage, and report margins on the original class distribution.
  2. [Abstract and Section I Contributions] The abstract claims 'over 10% in F1-score on two of the datasets' and the introduction claims improvements 'by over 10% in F1-score across three public datasets.' Table III shows only MulDiGraph exceeds 10% (10.24%); B4E is 8.85% and SPN is 6.49%. This is an internal factual inconsistency in a central claim. Please correct the wording to match the measured values, or present the numbers that justify the stated threshold.
  3. [Section III-A1, Eq. (2) and Table IV] The 'semantic' advantage is the load-bearing novelty of TxCLM, but the textualization in Eq. (2) simply prepends attribute names to numerical values (e.g., 'amount: 0.02', 'direction: 0'). The ablation w/o Language Model shows a 16.97% F1 drop on B4E, but this does not isolate whether the gain comes from the textual representation, the BERT architecture, or the contrastive pretraining. Please add an ablation that feeds the same BERT architecture the raw numerical sequences (without the attribute labels) or a matched numerical token sequence, so that the value of 'language semantics' is directly tested.
  4. [Section VI, Table III and Table IV] No error bars, standard deviations, or significance tests are reported for any of the main experiments, despite the text in Section VI-B referring to 'repeated experiments and statistical analysis.' Single-run results are insufficient to support the adverb 'consistently' and to quantify whether the 1-10% F1 gaps are stable across seeds and data subsamples. Please report mean and variance over at least 3-5 random seeds, and ideally paired tests against the strongest baselines.
minor comments (4)
  1. [Throughout] Typos and inconsistent notation: 'avaliable' (abstract), 'Backgroud' (Section II header), 'utiliz' (Section V), 'MuDigraph' vs 'MulDiGraph' (Table III text), and BAcc/BACC are used interchangeably.
  2. [Table VII] The row 'LABOR (Ours)' is misleading because LABOR is an existing sampler from reference [32], not a contribution of this paper. The contribution is the integration of LABOR into MAGAE. Please rename the row to 'LABOR [32]' or 'MAGAE+LABOR' in the table and text.
  3. [Section V, Table II] The B4E row says 597,258 nodes but the text in Section V-B4E mentions 594,038 normal accounts and 3,220 phishing accounts, which sums to 597,258. Please clarify the discrepancy between 'nodes' in Table II and the account counts in the text; also clarify whether the constructed transaction graph includes all transaction documents.
  4. [Section VI-D, Table VIII] The isolated-split protocol should be described more precisely: it selects all phishing accounts and twice as many benign accounts, extracts connected components, and splits components. State explicitly how the sampled benign accounts are chosen and why this does not bias the comparison in favor of the language model.

Circularity Check

1 steps flagged · score 2.0 of 10

Minor self-definitional validation of the token-contrastive loss; the central SOTA claim remains empirically grounded.

  1. self definitional [Section VI-C (Eq. 22, Fig. 2) vs. Section III-A2 (Eq. 4)]
    "To analyze the differences in token representations learned by TxCLM and other BERT models, we calculated the average self-similarity of a sequence ... The smaller the value of sim(x), the lower the similarity between token representations in sequence x, which means the token representations are more discriminative."

    The token-aware contrastive objective LTa in Eq. (4) is itself a cosine-similarity contrastive loss that explicitly pushes masked token representations away from all other token representations in the same sequence. The 'self-similarity' metric in Eq. (22) is the average pairwise cosine similarity of the same token representations. Thus the observation that TxCLM has lower self-similarity than MLM-only BERT is a direct consequence of the training objective, not an independent confirmation that the representations capture finer semantics. The non-circular evidence for LTa's value is the w/o LTa ablation (F1 0.7846 vs 0.8143), which is independent of the self-similarity metric.

full rationale

LMAE4Eth is an empirical method paper rather than a derivation paper. The central state-of-the-art claim is established by running 21 baselines under a common protocol on three datasets, not by deriving the target metric from its own assumptions. The only mildly circular element is the self-similarity analysis in Section VI-C: the metric of Eq. (22) measures the same pairwise cosine similarity that the LTa objective of Eq. (4) is explicitly designed to reduce, so the visualization is a sanity check of the loss rather than an independent validation. This is not load-bearing because the contribution of LTa is independently tested by the w/o LTa ablation. The paper's use of the authors' own SPN dataset (ref. [38]) is a self-citation, but it is a data-collection artifact (two-hop BFS from known phishing nodes) and the main results are also demonstrated on the external MulDiGraph and B4E datasets, so it does not make the evaluation circular. The random node-split protocol and the fact that the isolated-component control in Table VIII covers only 7 of 21 baselines are legitimate generalization/leakage concerns, but they are correctness risks rather than definitional circularity. Overall circularity is minimal.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The framework rests on a set of domain assumptions about pre-training and evaluation. The free parameters are the hyperparameters tuned on the same datasets (mask ratios, temperature, scaling factor, fanout) plus an unspecified embedding dimension. No new physical or conceptual entities are introduced; the learnable tokens in the fusion network are model parameters. The axioms are standard ML assumptions plus one paper-specific modeling choice: excluding addresses from the text to avoid tokenizer issues.

free parameters (6)
  • TxCLM masking ratio = 0.15 (MulDiGraph, SPN); 0.30 (B4E)
    Figure 4a shows optimal mask ratio depends on dataset; sensitivity tuned on same datasets.
  • TxCLM temperature tau = 0.1
    Figure 4b shows best temperature around 0.1 across datasets.
  • MAGAE node masking ratio = 0.5 to 0.8
    Figure 4c shows best performance in this range.
  • MAGAE SCE scaling factor gamma = 3 to 4
    Figure 4d shows higher gamma (up to 4) improves F1.
  • LABOR fanout = dataset-dependent, e.g., 10:10 generally
    Figure 5 shows best fanout not fixed; 10:10 is a middle point used in Table VII.
  • Embedding dimensions dLM, dh, dNode = not specified
    The paper never states hidden sizes, number of layers, or aggregate token counts; these are hand-chosen.
assumptions (5)
  • domain assumption BERT-style masked language modeling on textualized transaction sequences learns semantic features useful for fraud detection.
    The entire TxCLM design assumes MLM pretraining on 'amount: v, direction: d, timestamp: t' strings yields meaningful token embeddings; invoked throughout Section III-A.
  • domain assumption GraphMAE-style node feature reconstruction learns node representations that transfer to node classification.
    The MAGAE uses reconstruction of masked expert features to produce account embeddings; based on prior work (GraphMAE), assumed to hold on Ethereum graphs.
  • domain assumption The random node split with a shared graph does not introduce prohibitive label leakage.
    Main results use a 7:1:2 random node split on a single transaction graph, so test nodes are connected to training nodes; the paper assumes this is acceptable, though it also provides an isolated-component split as a check.
  • domain assumption Centrality and statistical expert features computed from the full graph are legitimate inputs for pretraining and do not themselves encode labels.
    Table I features include betweenness, eigenvector, Katz centrality, etc., computed over the whole graph used for pretraining and evaluation.
  • ad hoc to paper The textualization excludes addresses, so the language model does not memorize account identity.
    The authors deliberately drop Ethereum addresses from the text (Section III-A1) to avoid tokenizer issues; this is a modeling choice specific to this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LMAE4Eth: Generalizable and Robust Ethereum Fraud Detection by Exploring Transaction Semantics and Masked Graph Embedding." pith.science (2026). https://pith.science/paper/SN5Z7Y3L

@misc{pith2026250903939,
  author       = {Pith},
  title        = {Pith review of: LMAE4Eth: Generalizable and Robust Ethereum Fraud Detection by Exploring Transaction Semantics and Masked Graph Embedding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SN5Z7Y3L}},
  note         = {Machine review of arXiv:2509.03939}
}
read the original abstract

Current Ethereum fraud detection methods rely on context-independent, numerical transaction sequences, failing to capture semantic of account transactions. Furthermore, the pervasive homogeneity in Ethereum transaction records renders it challenging to learn discriminative account embeddings. Moreover, current self-supervised graph learning methods primarily learn node representations through graph reconstruction, resulting in suboptimal performance for node-level tasks like fraud account detection, while these methods also encounter scalability challenges. To tackle these challenges, we propose LMAE4Eth, a multi-view learning framework that fuses transaction semantics, masked graph embedding, and expert knowledge. We first propose a transaction-token contrastive language model (TxCLM) that transforms context-independent numerical transaction records into logically cohesive linguistic representations. To clearly characterize the semantic differences between accounts, we also use a token-aware contrastive learning pre-training objective together with the masked transaction model pre-training objective, learns high-expressive account representations. We then propose a masked account graph autoencoder (MAGAE) using generative self-supervised learning, which achieves superior node-level account detection by focusing on reconstructing account node features. To enable MAGAE to scale for large-scale training, we propose to integrate layer-neighbor sampling into the graph, which reduces the number of sampled vertices by several times without compromising training quality. Finally, using a cross-attention fusion network, we unify the embeddings of TxCLM and MAGAE to leverage the benefits of both. We evaluate our method against 21 baseline approaches on three datasets. Experimental results show that our method outperforms the best baseline by over 10% in F1-score on two of the datasets.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

55 extracted references · 43 canonical work pages

  1. [1]

    Artificial intelligence for cybersecurity: Literature review and future research directions,

    R. Kaur, D. Gabrijel ˇciˇc, and T. Klobu ˇcar, “Artificial intelligence for cybersecurity: Literature review and future research directions,” Information Fusion , vol. 97, p. 101804, 2023. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S1566253523001136

    work page 2023

  2. [2]

    Efficient and dos-resistant consensus for permissioned blockchains,

    X. Chen, S. Zhao, J. Qi, J. Jiang, H. Song, C. Wang, T. On Li, T. Hubert Chan, F. Zhang, X. Luo et al. , “Efficient and dos-resistant consensus for permissioned blockchains,” ACM SIGMETRICS Perfor- mance Evaluation Review , vol. 49, no. 3, pp. 61–62, 2022

    work page 2022

  3. [3]

    Blockchain challenges and opportunities: A survey,

    Z. Zheng, S. Xie, H.-N. Dai, X. Chen, and H. Wang, “Blockchain challenges and opportunities: A survey,” International journal of web and grid services , vol. 14, no. 4, pp. 352–375, 2018

    work page 2018

  4. [4]

    Understanding security issues in the dao governance process,

    J. Ma, M. Jiang, J. Jiang, X. Luo, Y . Hu, Y . Zhou, Q. Wang, and F. Zhang, “Understanding security issues in the dao governance process,” IEEE Transactions on Software Engineering , 2025

    work page 2025

  5. [5]

    Natural language processing in finance: A survey,

    K. Du, Y . Zhao, R. Mao, F. Xing, and E. Cambria, “Natural language processing in finance: A survey,” Information Fusion , vol. 115, p. 102755, 2025. [Online]. Available: https://www.sciencedirect. com/science/article/pii/S1566253524005335

    work page 2025

  6. [6]

    The Chainalysis 2024 Crypto Crime Report,

    Chainalysis, “The Chainalysis 2024 Crypto Crime Report,” 2024. [Online]. Available: https://go.chainalysis.com/crypto-crime-2024.html

    work page 2024

  7. [7]

    Lessons learned from blockchain applications of trusted execution environments and implications for future research,

    K. Rabimba, L. Xu, L. Chen, F. Zhang, Z. Gao, and W. Shi, “Lessons learned from blockchain applications of trusted execution environments and implications for future research,” in Proceedings of the 10th Inter- national Workshop on Hardware and Architectural Support for Security and Privacy, 2021, pp. 1–8

    work page 2021

  8. [8]

    Ttagn: Temporal transaction aggregation graph network for ethereum phishing scams detection,

    S. Li, G. Gou, C. Liu, C. Hou, Z. Li, and G. Xiong, “Ttagn: Temporal transaction aggregation graph network for ethereum phishing scams detection,” in Proceedings of the ACM Web Conference 2022 , 2022, pp. 661–669

    work page 2022

Show all 55 references
  1. [9]

    Graph neural network based phishing account detection in ethereum,

    S. Ratra, M. Ghosh, N. Baliyan, J. Rashmitha Mohan, and S. Singh, “Graph neural network based phishing account detection in ethereum,” The Computer Journal , p. bxae079, 2024

    2024

  2. [10]

    Ethereum phishing scams detection: A survey,

    Z. Zhang, M. Devaraj, X. Bai, and H. Lu, “Ethereum phishing scams detection: A survey,” in 2024 International Conference on Artificial Intelligence and Digital Technology (ICAIDT). IEEE, 2024, pp. 70–75

    2024

  3. [11]

    High-frequency trading on decentralized on-chain exchanges,

    L. Zhou, K. Qin, C. F. Torres, D. V . Le, and A. Gervais, “High-frequency trading on decentralized on-chain exchanges,” in 2021 IEEE Symposium on Security and Privacy (SP) . IEEE, 2021, pp. 428–445

    2021

  4. [12]

    Quantifying blockchain extractable value: How dark is the forest?

    K. Qin, L. Zhou, and A. Gervais, “Quantifying blockchain extractable value: How dark is the forest?” in 2022 IEEE Symposium on Security and Privacy (SP) . IEEE, 2022, pp. 198–214

    2022

  5. [13]

    node2vec: Scalable feature learning for networks,

    A. Grover and J. Leskovec, “node2vec: Scalable feature learning for networks,” in Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining , 2016, pp. 855– 864

    2016

  6. [14]

    Graph neural network for ethereum fraud detection,

    R. Tan, Q. Tan, P. Zhang, and Z. Li, “Graph neural network for ethereum fraud detection,” in 2021 IEEE International Conference on Big Knowledge (ICBK) , 2021, pp. 78–85

    2021

  7. [15]

    Detecting phishing scams on ethereum based on transaction records,

    Q. Yuan, B. Huang, J. Zhang, J. Wu, H. Zhang, and X. Zhang, “Detecting phishing scams on ethereum based on transaction records,” in2020 IEEE International Symposium on Circuits and Systems (ISCAS) , 2020, pp. 1– 5. 15

    2020

  8. [16]

    Who are the phishers? phishing scam detection on ethereum via network embedding,

    J. Wu, Q. Yuan, D. Lin, W. You, W. Chen, C. Chen, and Z. Zheng, “Who are the phishers? phishing scam detection on ethereum via network embedding,” IEEE Transactions on Systems, Man, and Cybernetics: Systems, vol. 52, no. 2, pp. 1156–1166, 2020

    2020

  9. [17]

    Deepwalk: Online learning of social representations,

    B. Perozzi, R. Al-Rfou, and S. Skiena, “Deepwalk: Online learning of social representations,” in Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining , 2014, pp. 701–710

    2014

  10. [18]

    Phishing fraud detection on ethereum using graph neural network,

    P. Li, Y . Xie, X. Xu, J. Zhou, and Q. Xuan, “Phishing fraud detection on ethereum using graph neural network,” in International Conference on Blockchain and Trustworthy Systems . Springer, 2022, pp. 362–375

    2022

  11. [19]

    Graph neural network for ethereum fraud detection,

    R. Tan, Q. Tan, P. Zhang, and Z. Li, “Graph neural network for ethereum fraud detection,” in 2021 IEEE international conference on big knowledge (ICBK) . IEEE, 2021, pp. 78–85

    2021

  12. [21]

    Blockchain phishing scam detection via multi-channel graph classification,

    D. Zhang, J. Chen, and X. Lu, “Blockchain phishing scam detection via multi-channel graph classification,” in Blockchain and Trustworthy Systems: Third International Conference, BlockSys 2021, Guangzhou, China, August 5–6, 2021, Revised Selected Papers 3 . Springer, 2021, pp. 241–256

    2021

  13. [22]

    Tsgn: Transaction subgraph networks assisting phishing detection in ethereum,

    J. Wang, P. Chen, X. Xu, J. Wu, M. Shen, Q. Xuan, and X. Yang, “Tsgn: Transaction subgraph networks assisting phishing detection in ethereum,” IEEE Transactions on Dependable and Secure Computing , 2025

    2025

  14. [23]

    Grabphisher: Phishing scams detection in ethereum via temporally evolving gnns,

    J. Zhang, H. Sui, X. Sun, C. Ge, L. Zhou, and W. Susilo, “Grabphisher: Phishing scams detection in ethereum via temporally evolving gnns,” IEEE Transactions on Services Computing , pp. 1–15, 2024

    2024

  15. [24]

    Bert4eth: A pre- trained transformer for ethereum fraud detection,

    S. Hu, Z. Zhang, B. Luo, S. Lu, B. He, and L. Liu, “Bert4eth: A pre- trained transformer for ethereum fraud detection,” in Proceedings of the ACM Web Conference 2023 , 2023, pp. 2189–2197

    2023

  16. [25]

    Zipzap: Efficient training of language models for large-scale fraud detection on blockchain,

    S. Hu, T. Huang, K.-H. Chow, W. Wei, Y . Wu, and L. Liu, “Zipzap: Efficient training of language models for large-scale fraud detection on blockchain,” in Proceedings of the ACM on Web Conference 2024, ser. WWW ’24. New York, NY , USA: Association for Computing Machinery, 2024...

    2024

  17. [26]

    Variational graph auto-encoders,

    T. N. Kipf and M. Welling, “Variational graph auto-encoders,” arXiv preprint arXiv:1611.07308, 2016

    2016 arXiv

  18. [27]

    Co-embedding attributed networks,

    Z. Meng, S. Liang, H. Bao, and X. Zhang, “Co-embedding attributed networks,” in Proceedings of the twelfth ACM international conference on web search and data mining , 2019, pp. 393–401

    2019

  19. [28]

    Variational graph auto-encoder based inductive learning method for semi-supervised clas- sification,

    H. Yang, Z. Yu, Q. Kong, W. Liu, and W. Mao, “Variational graph auto-encoder based inductive learning method for semi-supervised clas- sification,” arXiv preprint arXiv:2403.17500 , 2024

    2024 arXiv

  20. [29]

    Graphmae: Self-supervised masked graph autoencoders,

    Z. Hou, X. Liu, Y . Cen, Y . Dong, H. Yang, C. Wang, and J. Tang, “Graphmae: Self-supervised masked graph autoencoders,” in Proceed- ings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , 2022, pp. 594–604

    2022

  21. [30]

    Revisiting and benchmarking graph autoencoders: A con- trastive learning perspective,

    J. Li, R. Wu, Y . Zhu, H. Zhang, X. Jin, G. Zhang, Z. Zhu, Z. Zheng, and L. Chen, “Revisiting and benchmarking graph autoencoders: A con- trastive learning perspective,” arXiv preprint arXiv:2410.10241 , 2024

    2024 arXiv

  22. [31]

    S2gae: Self-supervised graph autoencoders are generalizable learners with graph masking,

    Q. Tan, N. Liu, X. Huang, S.-H. Choi, L. Li, R. Chen, and X. Hu, “S2gae: Self-supervised graph autoencoders are generalizable learners with graph masking,” in Proceedings of the sixteenth ACM international conference on web search and data mining , 2023, pp. 787–795

    2023

  23. [32]

    Layer-neighbor sampling—defusing neighborhood explosion in gnns,

    M. F. Balin and Ü. Çatalyürek, “Layer-neighbor sampling—defusing neighborhood explosion in gnns,” Advances in Neural Information Processing Systems, vol. 36, pp. 25 819–25 836, 2023

    2023

  24. [33]

    Attention bottlenecks for multimodal fusion,

    A. Nagrani, S. Yang, A. Arnab, A. Jansen, C. Schmid, and C. Sun, “Attention bottlenecks for multimodal fusion,” Advances in neural information processing systems , vol. 34, pp. 14 200–14 213, 2021

    2021

  25. [34]

    Unveiling the power of audio-visual early fusion transformers with dense interactions through masked modeling,

    S. Mo and P. Morgado, “Unveiling the power of audio-visual early fusion transformers with dense interactions through masked modeling,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 27 186–27 196

    2024

  26. [35]

    Attention is all you need,

    A. Vaswani, “Attention is all you need,” arXiv preprint arXiv:1706.03762, 2017

    2017 arXiv

  27. [36]

    Xblock-eth: Extracting and exploring blockchain data from ethereum,

    P. Zheng, Z. Zheng, J. Wu, and H.-N. Dai, “Xblock-eth: Extracting and exploring blockchain data from ethereum,” IEEE Open Journal of the Computer Society, vol. 1, pp. 95–106, 2020

    2020

  28. [37]

    XBLOCK Blockchain Datasets: InPlusLab ethereum phishing detection datasets,

    L. Chen, J. Peng, Y . Liu, J. Li, F. Xie, and Z. Zheng, “XBLOCK Blockchain Datasets: InPlusLab ethereum phishing detection datasets,” http://xblock.pro/ethereum/, 2019

    2019

  29. [38]

    Ethereum fraud detection via joint transaction language model and graph representation learning,

    Y . Jia, Y . Wang, J. Sun, Y . Liu, Z. Sheng, and Y . Tian, “Ethereum fraud detection via joint transaction language model and graph representation learning,” arXiv preprint arXiv:2409.07494 , 2024

    2024 arXiv

  30. [39]

    role2vec: Role-based network embeddings,

    N. K. Ahmed, R. A. Rossi, J. B. Lee, T. L. Willke, R. Zhou, X. Kong, and H. Eldardiry, “role2vec: Role-based network embeddings,” Proc. DLG KDD, pp. 1–7, 2019

    2019

  31. [40]

    Exploring the limits of transfer learning with a unified text-to-text transformer,

    C. Raffel, N. Shazeer, A. Roberts, K. Lee, S. Narang, M. Matena, Y . Zhou, W. Li, and P. J. Liu, “Exploring the limits of transfer learning with a unified text-to-text transformer,” Journal of machine learning research, vol. 21, no. 140, pp. 1–67, 2020

    2020

  32. [41]

    Longformer: The long- document transformer,

    I. Beltagy, M. E. Peters, and A. Cohan, “Longformer: The long- document transformer,” arXiv preprint arXiv:2004.05150 , 2020

    2004 arXiv

  33. [42]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” arXiv preprint arXiv:1609.02907 , 2016

    2016 arXiv

  34. [43]

    Inductive representation learning on large graphs,

    W. Hamilton, Z. Ying, and J. Leskovec, “Inductive representation learning on large graphs,” Advances in neural information processing systems, vol. 30, 2017

    2017

  35. [44]

    Grand: Graph neural diffusion,

    B. Chamberlain, J. Rowbottom, M. I. Gorinova, M. Bronstein, S. Webb, and E. Rossi, “Grand: Graph neural diffusion,” in International confer- ence on machine learning . PMLR, 2021, pp. 1407–1418

    2021

  36. [45]

    Graph random neural networks for semi-supervised learning on graphs,

    W. Feng, J. Zhang, Y . Dong, Y . Han, H. Luan, Q. Xu, Q. Yang, E. Khar- lamov, and J. Tang, “Graph random neural networks for semi-supervised learning on graphs,” Advances in neural information processing systems, vol. 33, pp. 22 092–22 103, 2020

    2020

  37. [46]

    Hierarchical graph representation learning with differentiable pooling,

    Z. Ying, J. You, C. Morris, X. Ren, W. Hamilton, and J. Leskovec, “Hierarchical graph representation learning with differentiable pooling,” Advances in neural information processing systems , vol. 31, 2018

    2018

  38. [47]

    Unsupervised Univer- sal Self-Attention Network for Graph Classification,

    D. Q. Nguyen, T. D. Nguyen, and D. Phung, “Unsupervised Univer- sal Self-Attention Network for Graph Classification,” arXiv preprint arXiv:1909.11855, 2019

    1909 arXiv

  39. [48]

    Phishing detection on ethereum via learning representation of transaction subgraphs,

    Z. Yuan, Q. Yuan, and J. Wu, “Phishing detection on ethereum via learning representation of transaction subgraphs,” in Blockchain and Trustworthy Systems: Second International Conference, BlockSys 2020, Dali, China, August 6–7, 2020, Revised Selected Papers 2 . Springer, 2020,...

    2020

  40. [49]

    graph2vec: Learning distributed representations of graphs,

    A. Narayanan, M. Chandramohan, R. Venkatesan, L. Chen, Y . Liu, and S. Jaiswal, “graph2vec: Learning distributed representations of graphs,” arXiv preprint arXiv:1707.05005 , 2017

    2017 arXiv

  41. [50]

    Catalog: Exploiting joint temporal dependencies for enhanced phishing detection on ethereum,

    M. Ghosh, S. Srivastava, A. Upadhyaya, R. Halder, and J. Chandra, “Catalog: Exploiting joint temporal dependencies for enhanced phishing detection on ethereum,” in Proceedings of the ACM on Web Conference 2025, 2025, pp. 969–977

    2025

  42. [51]

    Graph attention auto-encoders,

    A. Salehi and H. Davulcu, “Graph attention auto-encoders,” arXiv preprint arXiv:1905.10715, 2019

    1905 arXiv

  43. [52]

    Roberta: A robustly optimized bert pretraining approach,

    Y . Liu, “Roberta: A robustly optimized bert pretraining approach,” arXiv preprint arXiv:1907.11692, vol. 364, 2019

    1907 arXiv

  44. [53]

    Albert: A lite bert for self-supervised learning of language representations,

    Z. Lan, “Albert: A lite bert for self-supervised learning of language representations,” arXiv preprint arXiv:1909.11942 , 2019

    1909 arXiv

  45. [54]

    Electra: Pre-training text encoders as discriminators rather than generators,

    K. Clark, “Electra: Pre-training text encoders as discriminators rather than generators,” arXiv preprint arXiv:2003.10555 , 2020

    2003 arXiv

  46. [55]

    Vgcn-bert: augmenting bert with graph embedding for text classification,

    Z. Lu, P. Du, and J.-Y . Nie, “Vgcn-bert: augmenting bert with graph embedding for text classification,” in Advances in Information Retrieval: 42nd European Conference on IR Research, ECIR 2020, Lisbon, Por- tugal, April 14–17, 2020, Proceedings, Part I 42 . Springer, 2020, pp...

    2020

  47. [56]

    Layer-dependent importance sampling for training deep and large graph convolutional networks,

    D. Zou, Z. Hu, Y . Wang, S. Jiang, Y . Sun, and Q. Gu, “Layer-dependent importance sampling for training deep and large graph convolutional networks,” Advances in neural information processing systems , vol. 32, 2019

    2019

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.