Pith. sign in

REVIEW 4 major objections 6 minor 39 references

Dynamic Feature Fusion: Combining Global Graph Structures and Local Semantics for Blockchain Fraud Detection

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a dynamic fusion of GCN graph embeddings and BERT text embeddings detects Ethereum fraud with an F1 of 94.71 percent on one benchmark, beating the best baseline by roughly 20 points.

desk verdict The central results are invalid because the ground-truth tag is embedded in the BERT input, so the reported 94.71 F1 is token copying, not learned fusion. read the letter →

arxiv 2501.02032 v1 pith:NF65IJN2 submitted 2025-01-03 cs.CR cs.AIcs.SE

Zhang Sheng , Liangliang Song , Yanbin Wang This is my paper · ORCID
classification cs.CRcs.AIcs.SE
keywords EthereumfrauddetectionmultimodalfusiondynamicgatinggraphneuralnetworksBERTtextembeddingsGCNphishingn-gramtimedifference
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

The paper claims that Ethereum fraud detection improves substantially when global transaction-network structure and local transaction-text semantics are fused through a dynamic gating mechanism, rather than used alone or merged with fixed weights. It proposes ETH-GBERT, pairing a graph convolutional network over the account transaction graph with a BERT encoder over per-account transaction text, and a Gumbel-softmax gate that chooses, for each account, between text-only, graph-enhanced-text, and weighted-combination embeddings. On three real Ethereum datasets the model reports F1 scores up to 94.71%, roughly 20 points above the best random-walk baseline. The authors frame this as evidence that structural and semantic signals carry complementary fraud indicators that an adaptive fusion can exploit.

What carries the argument

The carrying object is the dynamic feature fusion module: a gating network $G(x)$ that takes the concatenation of the BERT embedding $E_{\text{BERT}}$ and the GCN-enhanced embedding $E_{\text{GCN\text{-}Enhanced}}$ and emits three weights $g_1,g_2,g_3$ through a Gumbel-softmax with temperature $\tau$, yielding $E_{\text{Fused}} = g_1 E_{\text{BERT}} + g_2 E_{\text{GCN\text{-}Enhanced}} + g_3 (\alpha E_{\text{BERT}} + (1-\alpha)E_{\text{GCN\text{-}Enhanced}})$. At low temperature the gate approximates a one-hot choice while the straight-through estimator keeps the network differentiable, so the model can pick the cheaper single-modality path or the full weighted blend per input. The graph side uses an adjacency matrix whose edge weights are transaction values multiplied by a sum of $n$-gram time differences $\Delta T_n = T_i - T_{i-(n-1)}$, designed to encode short-burst phishing behavior into the network structure.

What would settle it

Retrain ETH-GBERT on the same three datasets with the tag field removed from the text corpus, keeping every other preprocessing and hyperparameter step identical, and compare F1 against the same baselines; if the gap collapses, the original improvement is carried by label leakage inside the text rather than by dynamic multimodal fusion.

Watch

Extended reading notes

Core claim

On its own terms, the central discovery is that the dynamic fusion of BERT text embeddings and GCN graph embeddings yields consistent, large improvements over unimodal and static-fusion baselines for Ethereum fraud detection: F1 of 94.71% on the Multigraph dataset versus 74.13% for the strongest baseline (Role2Vec), 86.16% on the Transaction Network dataset versus 71.39%, and 89.79% with 89.57% recall on the B4E dataset. The gating network generates instance-specific weights over three fusion strategies, letting the model rely more on text for some accounts and more on graph structure for others. The authors attribute the gains to this adaptability and conclude that multimodal fusion is a scalable direction for blockchain fraud detection.

Load-bearing premise

The load-bearing premise is that the transaction text passed to BERT contains only innocuous metadata; the paper's own text-generation step writes the account tag into the sentence (for example 'AccountA : tag = 1, Value=5.06854256'), so if those inputs are what BERT sees, the reported superiority may largely measure the model reading a label token rather than learning to fuse graph structure and semantics.

Editorial extensions

If this is right

  • If the central claim is correct, adaptive graph-text fusion is a practical recipe for Ethereum phishing detection, adding roughly 20 F1 points over the best random-walk baseline on the Multigraph dataset.
  • Dynamic gating outperforms static weighted combination on all three datasets (for example, +2.28 F1 on Multigraph), identifying per-instance modality weighting as the active ingredient rather than mere feature concatenation.
  • BERT alone reaches 90.10 F1 on Multigraph, so text semantics carry much of the signal; the fusion gain is largest where text alone is weakest, such as the Transaction Network and B4E datasets.
  • Performance peaks at a balanced 5:5 normal-to-fraud ratio (94.71 F1 on Multigraph) and falls to the high 70s or low 80s at extreme imbalance, indicating that deployment must account for base-rate shifts.
  • The n-gram time-difference edge weighting provides a concrete way to fold temporal burst behavior into a static adjacency matrix, potentially benefiting other graph-based transaction models.

Reading between the lines

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

  • A reader building on this work should first remove the tag token from the text input and re-measure the gap: the paper's own pipeline places 'tag = 1' directly in the corpus, so genuine semantic fusion is currently confounded with label-reading.
  • The same Gumbel-softmax gating idea could be lifted to other graph-plus-text problems, such as anti-money-laundering or social-network abuse detection, wherever one modality is more reliable per instance.
  • The paper does not ablate the n-gram time-difference weighting on its own, so its contribution to the reported gains could be tested independently by replacing that adjacency weight with plain transaction counts.
  • The balanced-ratio experiments hint that real-world deployment value depends on the operating point: at realistic phishing base rates the reported F1 drops below 80 on some datasets.
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 / 6 minor

Summary. The manuscript proposes ETH-GBERT, a multimodal model for Ethereum fraud detection that combines GCN-based global graph representations with BERT-based local semantic representations of transaction text, using a Gumbel-Softmax gating network to fuse the two modalities. The authors describe a data processing pipeline with n-gram temporal features, adjacency-matrix construction, and text generation, and report experiments on three Ethereum datasets (Multigraph, Transaction Network, and B4E) with claimed state-of-the-art F1 scores of 94.71, 86.16, and 89.79, respectively. The paper also includes ablations over unimodal and static-fusion variants and a sensitivity analysis across normal-to-fraud ratios.

Significance. If the empirical claims were valid, the paper would provide a practical demonstration that dynamic fusion of graph structure and transaction text improves Ethereum fraud detection, and the data processing pipeline is described in enough detail to be a useful reference. The gating mechanism, however, is not new; it is directly adopted from DynMM. More importantly, the evaluation has a fundamental label-leakage problem: the text input to BERT contains the exact ground-truth tag token, so the reported results do not support the claimed effectiveness of the fusion mechanism. The stress-test concern is confirmed by the manuscript text, and the central SOTA claims in the abstract, introduction, and Section VI must be disregarded until a leak-free evaluation is performed.

major comments (4)
  1. [Section III-A3 and III-A4] The text corpus is constructed as "AccountA : tag = 1, Value=...", where "tag = 1" is precisely the account-level ground-truth label defined in the preceding paragraph (an account is fraudulent whenever any transaction in it has tag = 1). Section III-A4 then extracts the same label y from this corpus for supervision. The BERT input therefore contains the answer token that the model is trained to predict, so the model can achieve high F1 by copying the tag token rather than by learning from graph structure or transaction semantics. This invalidates the central empirical claim of the paper, including the 94.71 F1 on Multigraph and all comparisons in Tables I and II.
  2. [Section VI-A and Table I] The baseline comparisons are unfair under the current input construction: graph-only baselines such as DeepWalk, Role2Vec, GCN, and GAT never see the label token in their feature vectors, so the reported 20.58-point improvement over Role2Vec on Multigraph is forced by an information asymmetry rather than by the proposed fusion mechanism. Likewise, the BERT-only ablation in Table II is inflated by the same leak, and the incremental gain of ETH-GBERT over BERT-only (4.61 points on Multigraph) does not demonstrate that dynamic fusion adds any value.
  3. [Section VI and Table II] No standard deviations, confidence intervals, or significance tests are reported for any of the F1, recall, or precision values. The differences that motivate the dynamic-fusion claim are small (for example, +2.28 on Multigraph for ETH-GBERT over Weighted Combination), and without repeated runs or error bars the reader cannot assess whether these differences are stable or are artifacts of a single run. Table III presents a single number per ratio setting and likewise has no statistical grounding.
  4. [Section III-A2] The adjacency-matrix construction is described as creating a dense n x n zero matrix, with n equal to the number of unique accounts. For the Multigraph dataset, Section IV-A reports 2,973,489 nodes, so a dense float matrix would require roughly 2.97e6^2 x 8 bytes, i.e., tens of terabytes of memory. The paper does not describe a sparse representation or any out-of-core handling, so the graph construction step as written is not feasible on the datasets claimed, which undermines the reproducibility of the experiments.
minor comments (6)
  1. [Section IV-A1] "Ethernet" appears where "Ethereum" is intended in the description of the phishing transaction network.
  2. [Section IV-A3] The sentence "This dataset is able to capture multi-hop relationships between trading accounts and is is suitable for phishing account detection" contains a duplicated "is".
  3. [Table I and Section IV-B] The baseline name is written inconsistently as "GSAGE" in Table I and "GraphSAGE" in the text; this should be standardized.
  4. [Section III-A2] The summation in the transaction-weight formula is rendered as "NX n=1"; proper summation notation and a definition of N are needed for clarity.
  5. [Table I] For ETH-GBERT on the Multigraph dataset, F1, recall, and precision are all reported as exactly 94.71; given that the text does not state a rounding convention, this exact equality across all three metrics looks unusual and should be clarified.
  6. [Abstract and Introduction] The repository URL is incomplete, ending at "Dynamic_Feature"; a working link or an explicit availability statement is needed for reproducibility.

Circularity Check

1 steps flagged · score 9.0 of 10

The SOTA claim is undermined by label leakage: Section III-A3 places the ground-truth tag token in the BERT text input, and Section III-A4 extracts the label y from that same corpus, so the reported F1 scores largely measure token copying.

  1. self definitional [Section III-A3 'Text Transaction Data Generation' and Section III-A4 'Text Data Cleaning']
    "An account is labelled as fraudulent whenever there is a transaction in the account with tag = 1... This tag is given to the first transaction record of the account... The key fields of each transaction (e.g., label tag, transaction value, etc.) are combined... For example, a transaction record might be converted into the following form: AccountA : tag = 1, Value= 5.06854256, . . ."

    The text feature 'tag' is defined to be exactly the account-level ground-truth label: the paragraph before the example says an account gets tag=1 whenever any transaction in it is fraudulent, and Section III-A4 then feeds the corpus to BERT while extracting 'the account label (y)' from the same text. Therefore the BERT module's input already contains the answer token, and the classifier's output can be driven almost entirely by attending to 'tag = 1'. The reported F1 scores and the claimed 20.58-point improvement over graph-only baselines are thus forced by input construction rather than by learned dynamic fusion of graphs and semantics.

full rationale

The paper's central empirical claim is not self-contained against a clean evaluation. The load-bearing premise is that the BERT text input contains only innocuous transaction metadata, but the paper's own data-generation steps contradict this: Section III-A3 deletes addresses and timestamps, retains 'transaction Value and label', assigns the account tag = 1 as the label, and formats each record as 'AccountA : tag = 1, Value= ...'. Section III-A4 then extracts both the corpus and the supervision signal y from that same text. Hence the ETH-GBERT BERT branch sees the ground-truth label as a token, so the 94.71/86.16/89.79 F1 scores in Table I are consistent with copying the tag token, not with the proposed multimodal fusion. The comparison is also unfair to graph-only baselines, whose node features never contain the label. The dynamic-fusion ablation in Table II does not repair the leak because the BERT-only and combined variants all share the same contaminated input. This is not a matter of author intent or self-citation; it is an explicit construction in the paper's own methodology that makes the main prediction equivalent to reading its own input. No independent benchmark or code check is provided to override this. Accordingly, the circularity score is high, though not maximal because a non-leaked rerun might still show some genuine fusion benefit.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The main underreported degrees of freedom are the n-gram weighting coefficients and the GCN architecture; the evaluation also silently assumes the label is not in the text input, which the paper's own preprocessing contradicts.

free parameters (4)
  • alpha_n (n-gram time difference coefficients) = not reported
    In the edge weight formula w_k = Value_k * sum_{n=2..5} alpha_n * Delta_t_{n,k}, the coefficients alpha_n are not given and appear hand-set or learned; the graph structure and all downstream results depend on them.
  • alpha (fusion weight) = initialized to 0.5, learned
    In Section III-B4, EFusion = alpha * E_BERT + (1-alpha) * E_GCN_Enhanced; alpha is a learnable parameter, a free parameter fit during training.
  • Temperature tau (Gumbel-Softmax) = not specified
    The gating network uses temperature tau controlling fusion weight sharpness; no value is reported, and it affects the dynamic fusion behavior.
  • GCN hidden dimensions and layer count = not reported
    The GCN module's architecture (number of layers, hidden units, initial node features H(0)) is never specified; these choices are free parameters not auditable.
assumptions (3)
  • standard math GCN message passing (Kipf and Welling) is a valid way to aggregate neighbor information on the transaction graph.
    The paper invokes the standard GCN update in Section III-B2 without proving convergence or suitability for this domain.
  • domain assumption Pretrained BERT representations transfer to short transaction-text strings.
    Section III-B3 uses BERT without domain adaptation; no evidence is given that tokenized 'tag = 1, Value=...' text has semantic signal independent of the label.
  • domain assumption The transaction-level 'tag' field is the ground-truth label for the account.
    Section V-A3 derives account labels from transaction tags; this same tag is later embedded in the text input, creating the leakage that invalidates the evaluation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dynamic Feature Fusion: Combining Global Graph Structures and Local Semantics for Blockchain Fraud Detection." pith.science (2026). https://pith.science/paper/NF65IJN2

@misc{pith2026250102032,
  author       = {Pith},
  title        = {Pith review of: Dynamic Feature Fusion: Combining Global Graph Structures and Local Semantics for Blockchain Fraud Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NF65IJN2}},
  note         = {Machine review of arXiv:2501.02032}
}
read the original abstract

The advent of blockchain technology has facilitated the widespread adoption of smart contracts in the financial sector. However, current fraud detection methodologies exhibit limitations in capturing both global structural patterns within transaction networks and local semantic relationships embedded in transaction data. Most existing models focus on either structural information or semantic features individually, leading to suboptimal performance in detecting complex fraud patterns.In this paper, we propose a dynamic feature fusion model that combines graph-based representation learning and semantic feature extraction for blockchain fraud detection. Specifically, we construct global graph representations to model account relationships and extract local contextual features from transaction data. A dynamic multimodal fusion mechanism is introduced to adaptively integrate these features, enabling the model to capture both structural and semantic fraud patterns effectively. We further develop a comprehensive data processing pipeline, including graph construction, temporal feature enhancement, and text preprocessing. Experimental results on large-scale real-world blockchain datasets demonstrate that our method outperforms existing benchmarks across accuracy, F1 score, and recall metrics. This work highlights the importance of integrating structural relationships and semantic similarities for robust fraud detection and offers a scalable solution for securing blockchain systems.

Figures

Figures reproduced from arXiv: 2501.02032 by the authors.

Figure 1
Figure 1. Architecture of the Dynamic Feature Fusion Model for Blockchain Fraud Detection [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. 3D Comparative Analysis of Model Performance Metrics Across Datasets [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Heatmap of Model Performance Metrics Across Multiple Datasets [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Trends of F1 Score, Recall, and Precision Across Different Normal [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 24 canonical work pages

  1. [1]

    A. Pal, C. K. Tiwari, A. Behl, Blockchain technology in financial services: a comprehensive review of the literature, Journal of Global Operations and Strategic Sourcing 14 (1) (2021) 61–80

    2021

  2. [2]

    Bhowmik, T

    M. Bhowmik, T. S. S. Chandana, B. Rudra, Comparative study of machine learning algorithms for fraud detection in blockchain, in: 2021 5th international conference on computing methodologies and communication (ICCMC), IEEE, 2021, pp. 539–541

    work page 2021

  3. [3]

    Wenhua, F

    Z. Wenhua, F. Qamar, T.-A. N. Abdali, R. Hassan, S. T. A. Jafri, Q. N. Nguyen, Blockchain technology: security issues, healthcare applications, challenges and future trends, Electronics 12 (3) (2023) 546

    work page 2023

  4. [4]

    M. N. M. Bhutta, A. A. Khwaja, A. Nadeem, H. F. Ahmad, M. K. Khan, M. A. Hanif, H. Song, M. Alshamari, Y . Cao, A survey on blockchain technology: Evolution, architecture and security, Ieee Access 9 (2021) 61048–61073

    work page 2021

  5. [5]

    J.-Y . Lai, J. Wang, Y .-H. Chiu, Evaluating blockchain technology for reducing supply chain risks, Information Systems and e-Business Management 19 (4) (2021) 1089–1111

    work page 2021

  6. [6]

    Givargizov, Unstable financial and economic factors in the world and their influence on the development of blockchain technologies, Inter- national Humanitarian University Herald

    I. Givargizov, Unstable financial and economic factors in the world and their influence on the development of blockchain technologies, Inter- national Humanitarian University Herald. Economics and Management (01 2023). doi:10.32782/2413-2675/2023-55-11

    work page doi:10.32782/2413-2675/2023-55-11 2023

  7. [7]

    Zheng, J

    Z. Zheng, J. Su, J. Chen, D. Lo, Z. Zhong, M. Ye, Dappscan: building large-scale datasets for smart contract weaknesses in dapp projects, IEEE Transactions on Software Engineering (2024)

    2024

  8. [8]

    Review of blockchain application with Graph Neural Networks, Graph Convolutional Networks and Convolutional Neural Networks

    A. Ancelotti, C. Liason, Review of blockchain application with graph neural networks, graph convolutional networks and convolutional neural networks, arXiv preprint arXiv:2410.00875 (2024)

    work page Pith review arXiv 2024

Show all 39 references
  1. [9]

    K. S. Roy, M. E. Karim, P. B. Udas, Exploiting deep learning based classification model for detecting fraudulent schemes over ethereum blockchain, in: 2022 4th International Conference on Sustainable Tech- nologies for Industry 4.0 (STI), IEEE, 2022, pp. 1–6

    2022

  2. [10]

    J. Wu, J. Liu, Y . Zhao, Z. Zheng, Analysis of cryptocurrency transactions from a network perspective: An overview, Journal of Network and Computer Applications 190 (2021) 103139

    2021

  3. [11]

    Elmougy, L

    Y . Elmougy, L. Liu, Demystifying fraudulent transactions and illicit nodes in the bitcoin network for financial forensics, in: Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2023, pp. 3979–3990

    2023

  4. [12]

    S. Liu, B. Cui, W. Hou, A survey on blockchain abnormal transaction detection, in: International Conference on Blockchain and Trustworthy Systems, Springer, 2023, pp. 211–225

    2023

  5. [13]

    Osterrieder, S

    J. Osterrieder, S. Chan, J. Chu, Y . Zhang, B. H. Misheva, C. Mare, Enhancing security in blockchain networks: Anomalies, frauds, and advanced detection techniques, arXiv preprint arXiv:2402.11231 (2024)

    2024 arXiv

  6. [14]

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

    2021

  7. [15]

    Kanezashi, T

    H. Kanezashi, T. Suzumura, X. Liu, T. Hirofuchi, Ethereum fraud detection with heterogeneous graph neural networks, arXiv preprint arXiv:2203.12363 (2022)

    2022 arXiv

  8. [16]

    P. Li, Y . Xie, X. Xu, J. Zhou, Q. Xuan, Phishing fraud detection on ethereum using graph neural network, in: International Conference on Blockchain and Trustworthy Systems, Springer, 2022, pp. 362–375. 12

    2022

  9. [17]

    J. Wang, P. Chen, X. Xu, J. Wu, M. Shen, Q. Xuan, X. Yang, Tsgn: Transaction subgraph networks assisting phishing detection in ethereum, arXiv preprint arXiv:2208.12938 (2022)

    2022 arXiv

  10. [18]

    W. Hou, B. Cui, R. Li, Detecting phishing scams on ethereum using graph convolutional networks with conditional random field, in: 2022 IEEE 24th Int Conf on High Performance Computing & Communi- cations; 8th Int Conf on Data Science & Systems; 20th Int Conf on Smart City; 8th ...

    2022

  11. [19]

    T. Hu, X. Liu, T. Chen, X. Zhang, X. Huang, W. Niu, J. Lu, K. Zhou, Y . Liu, Transaction-based classification and detection approach for ethereum smart contract, Information Processing & Management 58 (2) (2021) 102462

    2021

  12. [20]

    Farrugia, J

    S. Farrugia, J. Ellul, G. Azzopardi, Detection of illicit accounts over the ethereum blockchain, Expert Systems with Applications 150 (2020) 113318

    2020

  13. [21]

    B. Pan, N. Stakhanova, Z. Zhu, Ethershield: Time-interval analysis for detection of malicious behavior on ethereum, ACM Transactions on Internet Technology 21 (1) (2024) 1–30

    2024

  14. [22]

    S. Li, G. Gou, C. Liu, C. Hou, Z. Li, 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

    2022

  15. [23]

    T. Wen, Y . Xiao, A. Wang, H. Wang, A novel hybrid feature fusion model for detecting phishing scam on ethereum using deep neural network, Expert Systems with Applications 211 (2023) 118463

    2023

  16. [24]

    Chen, S.-Z

    Z. Chen, S.-Z. Liu, J. Huang, Y .-H. Xiu, H. Zhang, H.-X. Long, Ethereum phishing scam detection based on data augmentation method and hybrid graph neural network model, Sensors 24 (12) (2024) 4022

    2024

  17. [25]

    W. Chen, Z. Zheng, E. C.-H. Ngai, P. Zheng, Y . Zhou, Exploiting blockchain data to detect smart ponzi schemes on ethereum, IEEE Access 7 (2019) 37575–37586

    2019

  18. [26]

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

    2016 arXiv

  19. [27]

    Devlin, Bert: Pre-training of deep bidirectional transformers for language understanding, arXiv preprint arXiv:1810.04805 (2018)

    J. Devlin, Bert: Pre-training of deep bidirectional transformers for language understanding, arXiv preprint arXiv:1810.04805 (2018)

    2018 arXiv

  20. [28]

    Vaswani, Attention is all you need, Advances in Neural Information Processing Systems (2017)

    A. Vaswani, Attention is all you need, Advances in Neural Information Processing Systems (2017)

    2017

  21. [29]

    Z. Xue, R. Marculescu, Dynamic multimodal fusion, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 2575–2584

    2023

  22. [30]

    L. Chen, J. Peng, Y . Liu, J. Li, F. Xie, Z. Zheng, Phishing scams de- tection in ethereum transaction network, ACM Transactions on Internet Technology (TOIT) 21 (1) (2020) 1–16

    2020

  23. [31]

    J. Wu, Q. Yuan, D. Lin, W. You, W. Chen, C. Chen, Z. Zheng, Who are the phishers? phishing scam detection on ethereum via network embedding, IEEE Transactions on Systems, Man, and Cybernetics: Systems 52 (2) (2020) 1156–1166

    2020

  24. [32]

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

    2023

  25. [33]

    Perozzi, R

    B. Perozzi, R. Al-Rfou, 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

  26. [34]

    Rozemberczki, R

    B. Rozemberczki, R. Sarkar, Fast sequence-based embedding with diffusion graphs, in: Complex Networks IX: Proceedings of the 9th Conference on Complex Networks CompleNet 2018 9, Springer, 2018, pp. 99–107

    2018

  27. [35]

    N. K. Ahmed, R. Rossi, J. B. Lee, T. L. Willke, R. Zhou, X. Kong, H. Eldardiry, Learning role-based graph embeddings, arXiv preprint arXiv:1802.02896 (2018)

    2018 arXiv

  28. [36]

    Béres, I

    F. Béres, I. A. Seres, A. A. Benczúr, M. Quintyne-Collins, Blockchain is watching you: Profiling and deanonymizing ethereum users, in: 2021 IEEE international conference on decentralized applications and infrastructures (DAPPS), IEEE, 2021, pp. 69–78

    2021

  29. [37]

    Hamilton, Z

    W. Hamilton, Z. Ying, J. Leskovec, Inductive representation learning on large graphs, Advances in neural information processing systems 30 (2017)

    2017

  30. [38]

    Veli ˇckovi´c, G

    P. Veli ˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Lio, Y . Bengio, Graph attention networks, arXiv preprint arXiv:1710.10903 (2017)

    2017 arXiv

  31. [39]

    LeCun, Y

    Y . LeCun, Y . Bengio, G. Hinton, Deep learning, nature 521 (7553) (2015) 436–444

    2015

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.