ATM-GAD: Adaptive Temporal Motif Graph Anomaly Detection for Financial Transaction Networks

Erkang Bao; Lin Song; Xiaoling Lv; Xinyue Wang; Zeyue Zhang

REVIEW 4 major objections 6 minor 35 references

ATM-GAD: Adaptive Temporal Motif Graph Anomaly Detection for Financial Transaction Networks

T0 review · 4 major / 6 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 For fraud detection, this paper claims, the winning move is time-stamped subgraph motifs inside a per-account learned window — a recipe that outperforms seven baselines on four real-world datasets.

desk verdict A reasonable temporal-motif anomaly detection architecture with an internally inconsistent evaluation; the per-node adaptive window idea is worth a look, but the reported SOTA numbers do not hold together. read the letter →

arxiv 2508.20829 v1 pith:GWQYOZMF submitted 2025-08-28 cs.LG

Zeyue Zhang , Lin Song , Erkang Bao , Xiaoling Lv , Xinyue Wang This is my paper · ORCID

classification cs.LG

keywords financialfrauddetectiontemporalmotifsgraphanomalyadaptivetimewindowsattentionmechanismtransactionnetworksneuralmoneylaundering

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

ATM-GAD makes a specific claim: fraud in transaction networks is best detected by two time-rooted signals that earlier graph detectors ignore — temporal motifs, small recurring subgraphs whose edge timestamps expose suspicious money flows as they unfold, and account-specific observation windows, because fraud often erupts in short bursts unique to each entity. The model therefore learns a separate differentiable time window for every node, extracts 3-node, 3-edge temporal motifs inside that window, and aggregates them with two attention layers: one that reasons within a single motif and one that combines evidence across motif types. If the reported experiments hold up, the payoff is a detector that reaches the best AUPRC in every setting across ETH, Elliptic++, Bitcoin Alpha, and Bitcoin Otc, beating seven baselines and flagging fraud patterns earlier methods miss. The load-bearing premise is the fairness of the empirical comparison: sampling, split construction, and hyperparameter details are deferred to an appendix that is missing from the preprint.

What carries the argument

The load-bearing object is the pair (δ_vi, M(vi)): a node-specific time window δ_vi = τ_max·σ(f_θ(z_vi)) learned from the node's own embedding, and the set of temporal-motif instances extracted within it. Each 3-node, 3-edge instance is weighed by recency, w = σ(δ_vi − (τ_max^u − t_vi)), which funnels gradients back into the window. IntraA then uses a supernode self-attention to compress each instance, and InterA uses SparseMAX attention to combine the 84 motif-type embeddings into a motif representation that is concatenated with the backbone GCN embedding and sent to an MLP classifier. The whole chain — GCN, window learner, both attention layers, classifier — is trained jointly with cross-e

What would settle it

Complete the missing appendix and independently re-run the documented protocol on all eight dataset variants: the central claim fails if ATM-GAD's AUPRC does not lead MotifGNN and COFD, or if the full model does not beat its best ablation variant on the same splits — on ETH-1000 the InterA-only variant already reports AUC 0.953 versus the full model's 0.944. A second check targets the mechanism itself: pit the learned window against a per-node best fixed window found by search; if search-tuned fixed windows match the adaptive ones, the differentiability claim is not doing the work.

Watch

Extended reading notes

Core claim

The paper's central claim is that static motifs and fixed global windows are the wrong instruments for financial fraud, and that replacing them with time-stamped motifs inside per-node learned windows measurably improves detection. Its discovery, on its own terms, is a fully differentiable pipeline: a Temporal Motif Extractor that condenses each account's history into time-stamped 3-node, 3-edge subgraphs; an Adaptive Time-Window Learner that sets each node's window as δ_vi = τ_max·σ(f_θ(z_vi)), so gradient descent itself chooses how far back to look; and dual attention — IntraA pooling the nodes within one motif instance via a supernode, InterA applying sparse attention across all 84 motif

Load-bearing premise

The central claim rests on the fairness of the empirical comparison: the preprint defers dataset sampling, split construction, and hyperparameter details to an appendix that is missing, and if those choices favor ATM-GAD or let information leak across the three train-test splits, the reported performance advantage could disappear.

Editorial extensions

If this is right

Transaction-graph detectors that still rely on static motifs have a concrete upgrade path: keep the 3-node subgraph structure but feed edge timestamps into a temporal extractor, since timing and ordering carry fraud signal topology alone misses.
Fixed global observation windows become a tunable liability rather than a default: the paper reports that larger fixed windows do not reliably improve results, while the learned per-node window keeps performance stable across temporal scopes.
Per-node window values and motif-count distributions are themselves informative: anomalous accounts concentrate motifs in short windows and learn narrow windows, so window statistics could join the feature set of non-graph fraud models.
Because the window learner is differentiable and the attention blocks sit on top of any GNN backbone, the temporal-motif layer can be retrofitted onto existing GCN- or GAT-based detectors.
End-to-end training means window selection is optimized against the fraud label, replacing grid-searched window hyperparameters with a learned quantity.

Reading between the lines

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

If per-node differentiable windows generalize beyond fraud, the same trick could sharpen other temporal-graph tasks, such as link prediction or event forecasting, where the lookback horizon is currently a global hyperparameter.
The observed separation between anomalous (concentrated, short-window) and normal (spread-out, diffuse-window) accounts suggests an unsupervised screening tool: even without labels, the slope of motif-count versus window length or the entropy of learned windows might rank suspicious accounts for human review.
The ablation results caution against reading the architecture as a uniform additive win: on the ETH-1000 variant the full model's reported AUC (0.944) is below the InterA-only variant's (0.953), so the joint benefit of the two attention levels should be validated per dataset rather than taken from the aggregate table.
A sharp test of the paper's core rationale: replace the learned window with the best fixed window chosen per node by search; if grid-searched fixed windows match the adaptive ones, then node-specificity rather than differentiability is carrying the gain.

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.

Desk Editor's Note

read the letter

Short version: the paper has a sensible architectural idea and a self-contradicting evaluation. The per-node learnable time window for temporal-motif extraction, plus intra/inter motif attention, is a reasonable combination that I don't think appears in the cited literature. That part is worth engaging. But the empirical claim of consistent SOTA is not supported by the numbers as reported.

What the paper does well: the problem framing is right. Fraud accounts act in bursts; global fixed windows are a real limitation. The Adaptive Time-Window Learner via sigmoid scaled by tau_max is simple and differentiable, and the instance weighting in Eq. (4) does let gradients flow to delta_vi. The two-level attention is standard but applied cleanly. Related work is cited appropriately. I do not see a circularity problem: the windows are learned outputs, not parameters tuned to the test set.

Where it falls apart: Table 1 and Table 3 report different AUPRC for the same model and dataset. ETH-200 gives 0.739 vs 0.871; ETH-1000 gives 0.844 vs 0.742. Worse, the ablation's “Only GCN” on ETH-200 gets 0.863 AUPRC, which would beat the full model's 0.739 in Table 1. That makes the headline “state-of-the-art on all settings” load-bearing but unstable. The paper says protocol details are in Appendix ??; all appendices and Algorithm ?? are placeholders. No code is released. So there is no way to tell which table is correct or whether the splits and subgraph construction were handled cleanly. The complexity bound O(|V| d^2) is asserted without derivation, and given per-node extraction over 84 motif types, that claim needs real support. Also, the adaptive-vs-fixed comparison in Table 2 only shows two datasets, and even within Table 1 the baselines win on AUC for some datasets, undercutting “consistently outperforms.”

Who this is for: someone working on temporal motif features for graph fraud detection will find the architecture worth thinking about. But this manuscript is not in a reviewable state. The fix is straightforward in principle: complete the appendices, reconcile the two tables, release code and seeds, and report the adaptive-window comparison on all datasets. Until then, the empirical claim cannot be evaluated.

Recommendation: I would not send this to referees in its current form. I would desk-reject as incomplete and invite a resubmission once the protocol details are filled in and the numbers agree.

Referee Report

4 major / 6 minor

Summary. The paper introduces ATM-GAD, a supervised GNN-based method for account-level fraud detection in financial transaction networks. The model learns a per-node adaptive time window via a sigmoid-parameterized network, extracts 3-node 3-edge temporal motifs within each node-specific window, and combines a GCN backbone with two attention modules (IntraA within a motif, InterA across motif types) to produce final node embeddings for binary classification. The authors report experiments on four datasets (ETH, Elliptic++, Bitcoin Alpha, Bitcoin Otc) with several subgraph variants, comparing against seven baselines and claiming state-of-the-art AUPRC in all settings. The proposed architecture is plausible and the idea of differentiable per-node time windows is interesting, but the current manuscript contains unresolved numerical inconsistencies between the main table and the ablation table, missing appendices and algorithm placeholders, and an unsubstantiated complexity bound. These issues directly affect the credibility of the central empirical claim.

Significance. If the empirical claims were fully supported, ATM-GAD would be a useful contribution: it combines temporal motifs, per-node adaptive windows, and multi-level attention in a way that is not present in the cited prior work, and the paper explicitly identifies a plausible failure mode of fixed-window temporal motif extraction. The method is coherent, and the idea of making the time window differentiable (Eq. 1 and Eq. 4) is a genuine point of interest. However, the paper's central claim is empirical, and the evidence as presented is not reliable: the main results table and the ablation table report conflicting numbers for the same model on the same datasets, no error bars or significance tests appear in the main body, and the experimental protocol is deferred to appendices that are placeholders. The manuscript therefore does not currently establish its headline result.

major comments (4)

[Table 1 vs. Table 3] The two tables report conflicting ATM-GAD results on the same ETH datasets. For ETH-200, Table 1 gives AUPRC 0.739 and AUC 0.911, while Table 3 gives AUPRC 0.871 and AUC 0.911. For ETH-1000, Table 1 gives AUPRC 0.844 and AUC 0.961, while Table 3 gives AUPRC 0.742 and AUC 0.944. These differences are far beyond rounding and cannot both describe the same configuration. This makes the Section 5.2.1 claim that ATM-GAD achieves state-of-the-art performance on all settings in terms of AUPRC unstable: a reader cannot tell which number is the true performance. Moreover, on ETH-200 the Table 3 'Only GCN' row has AUPRC 0.863, which is higher than the full ATM-GAD AUPRC of 0.739 in Table 1, contradicting the paper's own progressive-ablation logic. The authors must reconcile these numbers and specify exactly which splits, subgraph variants, and hyperparameters each table uses.
[§5.1 and all 'Appendix ??' references] Section 5.1 states that dataset sampling strategies, split construction, and hyperparameter details are presented in 'Appendix ??', and the text repeatedly refers to 'Appendix ??', 'Algorithm ??', and 'Table ??' (e.g., §1, §4, §4.3, Table 1 note, Figure captions). None of these appendices or algorithms exist in the submitted manuscript, and no code is released. Because the central claim is an empirical performance claim, the missing protocol is not a formatting issue: without the exact subgraph construction, train/test split procedure, label processing, and hyperparameter choices, the reported results cannot be reproduced or even fully interpreted. The authors need to supply the complete experimental details and, ideally, a code release.
[§4.3, Complexity analysis] The paper states that total complexity is dominated by O(|V| \bar d^2), where \bar d is average degree, and refers to an absent appendix. This bound is not justified by the described algorithm. The method extracts, for each node, all 3-node 3-edge temporal motifs within a node-specific window, with up to Max=84 motif types and potentially many instances per type; the time-window pooling in Eq. (4) also involves all instances. The complexity of temporal motif enumeration depends on the number of candidate edge triples and the window length, not simply on the final GCN convolution cost. The authors should either provide a rigorous derivation of the stated bound or replace it with an honest account of the extraction and attention costs. As written, the bound is unsubstantiated and likely misleading.
[§5.1 and Table 1] The main results are reported as averages over three independent train-test splits, but Table 1 contains no standard deviations, confidence intervals, or significance tests; the note says complete results with standard deviations are in an absent appendix. Given the small sizes of some subgraphs (e.g., ETH-200) and the likely class imbalance, the claimed margins over baselines may not be stable. The '1st count' row is also not defined: it appears to count something across AUC and AUPRC, but the counting rule is unclear. The authors should report per-split results and variability in the main text, or at minimum include the appendix with the fully populated table.

minor comments (6)

[Eq. (3)] The softmax formula for α_v is missing the normalization denominator: it should read α_v = exp(s_intra(h_v)) / Σ_{v'∈M_u} exp(s_intra(h_v')). As printed, the expression is malformed.
[Abstract and Section 5.2.1] The abstract claims ATM-GAD 'consistently outperforms seven strong anomaly-detection baselines' without stating the metric, while Section 5.2.1 narrows this to AUPRC. This should be clarified, since in Table 1 ATM-GAD does not achieve the best AUC on every dataset (e.g., Elli-1000 and Bitcoin Otc AUC values are lower than Random Forest).
[§4.1 and §4.2] The number Max=84 of 3-node 3-edge motif types is stated without derivation or citation. The standard count of connected directed 3-node, 3-edge isomorphism classes is much smaller, so the reader cannot verify the enumeration. A table or reference should be provided.
[Figure 2 caption] The caption contains stray text ('This is for caption: ...', '× × = ==') that appears to be leftover from figure assembly. This must be cleaned up.
[§5.2.3, Table 2] The phrase 'when δ_adp equals δ_fixed' is ambiguous, since δ_adp is a learned vector and δ_fixed is a scalar. The comparison would benefit from a precise statement of how the two settings are matched.
[References and related work] The related work is concise, but some recent temporal-motif anomaly detection works are mentioned only in passing. At minimum, the missing appendix references should be resolved; if any cited 'Appendix' content was intended to be part of the submission, it must be included.

Circularity Check

1 steps flagged · score 4.0 of 10

Adaptive time window is defined circularly via the final node embedding it determines; no other circular steps found.

self definitional [Section 4.1, Eq. (1); Section 4.2 'Final node embedding']
"δvi = τmax · σ (fθ(zvi )) (1) where zvi denotes the final node embeddings ... Given time window [tvi , tvi + δvi ] ... we extract all possible 3-node 3-edge temporal motifs involving vi. ... The temporal motif embedding for node vi is simply the sum ... ˜hvi = P κj βv,κj hvi,κj . ... zvi = [hvi ∥ ˜hvi]."

Equation (1) defines the adaptive window δvi as a function of zvi, explicitly called the final node embedding. But the final node embedding is defined later as the concatenation of the GCN embedding hvi and the motif embedding ˜hvi, where ˜hvi is computed from temporal motifs extracted using δvi in the window [tvi, tvi+δvi]. Hence δ is a function of z, and z is a function of δ; no fixed-point iteration, stop-gradient, or alternative input is specified. The 'adaptive per-node time window' is therefore self-referential by construction, rather than being an independently derived model input.

full rationale

The central claim is an empirical performance claim (SOTA AUPRC on four datasets), supported by supervised training and held-out evaluation; that claim does not reduce to the model's equations. No fitted parameter is renamed as a prediction, and there is no load-bearing self-citation chain: the temporal-motif definitions come from external prior work ([13,25]), and the supernode technique is credited to [35]. The only circular step I can exhibit is the self-referential definition of the adaptive time window: Eq. (1) computes δvi from the final embedding zvi, while zvi is later defined using motif embeddings extracted from windows of length δvi. As written, this is a cyclic dependency, although the empirical comparisons could still be valid if the implementation uses the initial GCN embedding or an unrolling not described. The Table 1 vs Table 3 AUPRC discrepancies and the missing Appendix ?? are reproducibility/correctness problems, not circularity, so they do not raise the circularity score further.

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

No new physical or conceptual entities are introduced; the supernode mechanism is borrowed from Yuan et al. [35]. The main fitted quantities are the per-node windows and standard trainable weights, plus hand-chosen structural choices (84 motif types, GCN depth/width).

free parameters (3)

Node-specific time window delta_vi = learned per node via f_theta(z_vi)
Central to the claimed advantage of adaptive windows over fixed ones; trained end-to-end with cross-entropy (Section 4.1, Eq. 1 and Eq. 4).
Number of motif types Max = 84
Hand-chosen enumeration of 3-node 3-edge directed temporal motifs; no derivation or appendix provided (Section 4.1).
GCN layers and hidden dimensions = 2-4 layers; dims {16, 32, 64}
Chosen by hand; sensitivity analysis in Figure 5 shows robustness but performance varies with these choices (Section 5.1).

assumptions (4)

domain assumption The 84 hand-enumerated 3-node 3-edge temporal motif types are sufficient to capture fraud-relevant patterns.
No justification beyond tractability; Section 4.1 'with Max=84 in our setting, see Appendix ??'.
domain assumption Transaction timestamps alone (without amounts or other edge features) carry enough signal for fraud detection; node features X are used in GCN but their content is never specified.
The method ignores edge weights/amounts except qualitatively in Figure 1; Section 3.1 defines X but experiments do not describe it.
ad hoc to paper Gradient-based optimization of delta through Eq. 4 yields meaningful per-node windows despite the hard threshold used in motif extraction.
Section 4.1 claims full differentiability via weighted pooling, but extraction uses a hard window cutoff, so the forward/backward mismatch is unexamined.
domain assumption Ground-truth labels in ETH, Elliptic++, Bitcoin Alpha, and Bitcoin Otc are reliable.
Benchmark labels are taken as correct; Section 5.1.

how reviews work

0 comments

Cite this review

Pith. "Pith review of ATM-GAD: Adaptive Temporal Motif Graph Anomaly Detection for Financial Transaction Networks." pith.science (2026). https://pith.science/paper/GWQYOZMF

@misc{pith2026250820829,
  author       = {Pith},
  title        = {Pith review of: ATM-GAD: Adaptive Temporal Motif Graph Anomaly Detection for Financial Transaction Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GWQYOZMF}},
  note         = {Machine review of arXiv:2508.20829}
}

read the original abstract

Financial fraud detection is essential to safeguard billions of dollars, yet the intertwined entities and fast-changing transaction behaviors in modern financial systems routinely defeat conventional machine learning models. Recent graph-based detectors make headway by representing transactions as networks, but they still overlook two fraud hallmarks rooted in time: (1) temporal motifs--recurring, telltale subgraphs that reveal suspicious money flows as they unfold--and (2) account-specific intervals of anomalous activity, when fraud surfaces only in short bursts unique to each entity. To exploit both signals, we introduce ATM-GAD, an adaptive graph neural network that leverages temporal motifs for financial anomaly detection. A Temporal Motif Extractor condenses each account's transaction history into the most informative motifs, preserving both topology and temporal patterns. These motifs are then analyzed by dual-attention blocks: IntraA reasons over interactions within a single motif, while InterA aggregates evidence across motifs to expose multi-step fraud schemes. In parallel, a differentiable Adaptive Time-Window Learner tailors the observation window for every node, allowing the model to focus precisely on the most revealing time slices. Experiments on four real-world datasets show that ATM-GAD consistently outperforms seven strong anomaly-detection baselines, uncovering fraud patterns missed by earlier methods.

Figures

Figures reproduced from arXiv: 2508.20829 by the authors.

**Figure 1.** Transactions of fraudulent account v0 over time from the Ethereum dataset[4]. During its early stage (dashed box), the account initiates a burst of large transfers; afterwards activity recedes to occasional lowvalue transactions. Even within the burst, the flow evolves from small transfers to a sudden spike of high-value ones and back again. Using the entire time span as a window dilutes the anomalous burst with… view at source ↗

**Figure 2.** The Overview of ATM-GAD. (a) Node embeddings are computed via GCN[ [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** (a) Distribution of temporal motif counts at varying time windows [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Distribution of the learned adaptive time window δada. We make an in-depth analysis of the proposed adaptive window selection when computing temporal motifs for each node. Firstly, we compare our learnable window δada with a conventional fixed window δfixed [PITH_FU…

**Figure 5.** Figure 5: Parameter sensitivity studies of ATM-GAD. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 27 canonical work pages

[1]

Blockchain: A Graph Primer

Cuneyt Gurcan Akcora, Yulia R Gel, and Murat Kantarcioglu. Blockchain: A graph primer. arXiv preprint arXiv:1708.08749, 2017

work page Pith review arXiv 2017
[2]

Data mining for credit card fraud: A comparative study

Siddhartha Bhattacharyya, Sanjeev Jha, Kurian Tharakunnel, and J Christopher Westland. Data mining for credit card fraud: A comparative study. Decision support systems, 50(3):602–613, 2011

work page 2011
[3]

Random forests

Leo Breiman. Random forests. Machine learning, 45:5–32, 2001

2001
[4]

XBLOCK Blockchain Datasets: InPlusLab ethereum phishing detection datasets

Liang Chen, Jiaying Peng, Yang Liu, Jintang Li, Fenfang Xie, and Zibin Zheng. XBLOCK Blockchain Datasets: InPlusLab ethereum phishing detection datasets. http://xblock.pro/ ethereum/, 2019

work page 2019
[5]

Xgboost: A scalable tree boosting system

Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, pages 785–794, 2016

2016
[6]

Understanding ethereum via graph analysis

Ting Chen, Zihao Li, Yuxiao Zhu, Jiachi Chen, Xiapu Luo, John Chi-Shing Lui, Xiaodong Lin, and Xiaosong Zhang. Understanding ethereum via graph analysis. ACM Transactions on Internet Technology (TOIT), 20(2):1–32, 2020

work page 2020
[7]

Motif graph neural network

Xuexin Chen, Ruichu Cai, Yuan Fang, Min Wu, Zijian Li, and Zhifeng Hao. Motif graph neural network. IEEE Transactions on Neural Networks and Learning Systems, 2023

work page 2023
[8]

Generating long sequences with sparse transformers

Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509, 2019

arXiv 1904

Show all 35 references

[9]

Graph anomaly detection via multi-scale contrastive learning networks with augmented view

Jingcan Duan, Siwei Wang, Pei Zhang, En Zhu, Jingtao Hu, Hu Jin, Yue Liu, and Zhibin Dong. Graph anomaly detection via multi-scale contrastive learning networks with augmented view. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 7459–7467, 2023

2023
[10]

Arise: Graph anomaly detection on attributed networks via substructure awareness

Jingcan Duan, Bin Xiao, Siwei Wang, Haifang Zhou, and Xinwang Liu. Arise: Graph anomaly detection on attributed networks via substructure awareness. IEEE transactions on neural networks and learning systems, 2023

2023
[11]

Provably powerful graph neural networks for directed multigraphs

Béni Egressy, Luc V on Niederhäusern, Jovan Blanuša, Erik Altman, Roger Wattenhofer, and Kubilay Atasu. Provably powerful graph neural networks for directed multigraphs. In Pro- ceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 11838–11846, 2024

2024
[12]

Demystifying fraudulent transactions and illicit nodes in the bitcoin network for financial forensics

Youssef Elmougy and Ling 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, pages 3979–3990, 2023

2023
[13]

Scalable motif counting for large-scale temporal graphs

Zhongqiang Gao, Chuanqi Cheng, Yanwei Yu, Lei Cao, Chao Huang, and Junyu Dong. Scalable motif counting for large-scale temporal graphs. In 2022 IEEE 38th International Conference on Data Engineering (ICDE), pages 2656–2668. IEEE, 2022

2022
[14]

Inductive representation learning on large graphs

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

2017
[15]

Samcl: Subgraph-aligned multiview contrastive learning for graph anomaly detection

Jingtao Hu, Bin Xiao, Hu Jin, Jingcan Duan, Siwei Wang, Zhao Lv, Siqi Wang, Xinwang Liu, and En Zhu. Samcl: Subgraph-aligned multiview contrastive learning for graph anomaly detection. IEEE Transactions on Neural Networks and Learning Systems, 2023. 10

2023
[16]

Collaborative fraud detection: How collaboration impacts fraud detection

Jinzhang Hu, Ruimin Hu, Zheng Wang, Dengshi Li, Junhang Wu, Lingfei Ren, Yilong Zang, Zijun Huang, and Mei Wang. Collaborative fraud detection: How collaboration impacts fraud detection. In Proceedings of the 31st ACM International Conference on Multimedia , pages 8891–8899, 2023

2023
[17]

Hybrid-order anomaly detection on attributed networks

Ling Huang, Ye Zhu, Yuefang Gao, Tuo Liu, Chao Chang, Caixing Liu, Yong Tang, and Chang- Dong Wang. Hybrid-order anomaly detection on attributed networks. IEEE Transactions on Knowledge and Data Engineering, 35(12):12249–12263, 2021

2021
[18]

Semi-supervised classification with graph convolutional networks

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

2016 arXiv
[19]

Data mining techniques for the detection of fraudulent financial statements

Efstathios Kirkos, Charalambos Spathis, and Yannis Manolopoulos. Data mining techniques for the detection of fraudulent financial statements. Expert systems with applications, 32(4):995– 1003, 2007

2007
[20]

Rev2: Fraudulent user prediction in rating platforms

Srijan Kumar, Bryan Hooi, Disha Makhija, Mohit Kumar, Christos Faloutsos, and VS Subrah- manian. Rev2: Fraudulent user prediction in rating platforms. In Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining, pages 333–341. ACM, 2018

2018
[21]

Edge weight prediction in weighted signed networks

Srijan Kumar, Francesca Spezzano, VS Subrahmanian, and Christos Faloutsos. Edge weight prediction in weighted signed networks. In Data Mining (ICDM), 2016 IEEE 16th International Conference on, pages 221–230. IEEE, 2016

2016
[22]

Slade: Detecting dynamic anomalies in edge streams without labels via self-supervised learning

Jongha Lee, Sunwoo Kim, and Kijung Shin. Slade: Detecting dynamic anomalies in edge streams without labels via self-supervised learning. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 1506–1517, 2024

2024
[23]

Temporal motifs for financial networks: A study on mercari, jpmc, and venmo platforms

Penghang Liu, Rupam Acharyya, Robert E Tillman, Shunya Kimura, Naoki Masuda, and Ahmet Erdem Sarıyüce. Temporal motifs for financial networks: A study on mercari, jpmc, and venmo platforms. arXiv preprint arXiv:2301.07791, 2023

2023 arXiv
[24]

Using motif transitions for temporal graph generation

Penghang Liu and Ahmet Erdem Sariyüce. Using motif transitions for temporal graph generation. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 1501–1511, 2023

2023
[25]

Motifs in temporal networks

Ashwin Paranjape, Austin R Benson, and Jure Leskovec. Motifs in temporal networks. In Proceedings of the tenth ACM international conference on web search and data mining, pages 601–610, 2017

2017
[26]

Pytorch: An imperative style, high-performance deep learning library

Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019

2019
[27]

Raphtory

Pometry. Raphtory. https://github.com/Pometry/Raphtory, 2025. Accessed: 2025-04- 30

2025
[28]

Scalable temporal motif densest subnetwork discovery

Ilie Sarpe, Fabio Vandin, and Aristides Gionis. Scalable temporal motif densest subnetwork discovery. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2536–2547, 2024

2024
[29]

Attention is all you need

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

2017
[30]

Graph attention networks

Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017

2017 arXiv
[31]

Financial default prediction via motif-preserving graph neural network with curriculum learning

Daixin Wang, Zhiqiang Zhang, Yeyu Zhao, Kai Huang, Yulin Kang, and Jun Zhou. Financial default prediction via motif-preserving graph neural network with curriculum learning. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2233–224...

2023
[32]

Mcogcn-motif high-order feature-guided embedding learning framework for social link prediction

Nan Xiang, Wenjing Yang, and Xindi Rao. Mcogcn-motif high-order feature-guided embedding learning framework for social link prediction. Scientific Reports, 14(1):29504, 2024

2024
[33]

Motif-consistent counterfactuals with adversarial refinement for graph-level anomaly detection

Chunjing Xiao, Shikang Pang, Wenxin Tai, Yanlong Huang, Goce Trajcevski, and Fan Zhou. Motif-consistent counterfactuals with adversarial refinement for graph-level anomaly detection. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages...

2024
[34]

Temporal-amount snapshot multigraph for ethereum transaction tracking

Yunyi Xie, Jie Jin, Jian Zhang, Shanqing Yu, and Qi Xuan. Temporal-amount snapshot multigraph for ethereum transaction tracking. In Blockchain and Trustworthy Systems: Third International Conference, BlockSys 2021, Guangzhou, China, August 5–6, 2021, Revised Selected Papers 3,...

2021
[35]

Motif-level anomaly detection in dynamic graphs

Zirui Yuan, Minglai Shao, and Qiben Yan. Motif-level anomaly detection in dynamic graphs. IEEE Transactions on Information Forensics and Security, 18:2870–2882, 2023. 12

2023

Pith tools

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

[1] [1]

Blockchain: A Graph Primer

Cuneyt Gurcan Akcora, Yulia R Gel, and Murat Kantarcioglu. Blockchain: A graph primer. arXiv preprint arXiv:1708.08749, 2017

work page Pith review arXiv 2017

[2] [2]

Data mining for credit card fraud: A comparative study

Siddhartha Bhattacharyya, Sanjeev Jha, Kurian Tharakunnel, and J Christopher Westland. Data mining for credit card fraud: A comparative study. Decision support systems, 50(3):602–613, 2011

work page 2011

[3] [3]

Random forests

Leo Breiman. Random forests. Machine learning, 45:5–32, 2001

2001

[4] [4]

XBLOCK Blockchain Datasets: InPlusLab ethereum phishing detection datasets

Liang Chen, Jiaying Peng, Yang Liu, Jintang Li, Fenfang Xie, and Zibin Zheng. XBLOCK Blockchain Datasets: InPlusLab ethereum phishing detection datasets. http://xblock.pro/ ethereum/, 2019

work page 2019

[5] [5]

Xgboost: A scalable tree boosting system

Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, pages 785–794, 2016

2016

[6] [6]

Understanding ethereum via graph analysis

Ting Chen, Zihao Li, Yuxiao Zhu, Jiachi Chen, Xiapu Luo, John Chi-Shing Lui, Xiaodong Lin, and Xiaosong Zhang. Understanding ethereum via graph analysis. ACM Transactions on Internet Technology (TOIT), 20(2):1–32, 2020

work page 2020

[7] [7]

Motif graph neural network

Xuexin Chen, Ruichu Cai, Yuan Fang, Min Wu, Zijian Li, and Zhifeng Hao. Motif graph neural network. IEEE Transactions on Neural Networks and Learning Systems, 2023

work page 2023

[8] [8]

Generating long sequences with sparse transformers

Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509, 2019

arXiv 1904