Pith. sign in

REVIEW 4 major objections 6 minor 56 references

Bridging Cognitive Neuroscience and Graph Intelligence: Hippocampus-Inspired Multi-View Hypergraph Learning for Web Finance Fraud

T0 review · 4 major / 6 minor · reviewed 2026-08-03 · deepseek-v4-flash

Pith's one-line read The paper claims that blending hippocampal-style conflict detection with novelty-weighted hypergraph message passing outperforms all 15 baselines on six web finance fraud datasets.

desk verdict The architecture is coherent and the method plausibly works, but the headline numbers are internally swapped: the abstract's 39.14% AP gain matches Table 2's F1 improvement, and its 9.74% F1 matches the actual AP gain. read the letter →

arxiv 2601.11073 v3 pith:A2QAUNP4 submitted 2026-01-16 cs.LG cs.AI

classification cs.LGcs.AI
keywords frauddetectiongraphneuralnetworkhypergraphlearningmulti-viewlong-taileddistributioncamouflagedhippocampus-inspiredmodelmessagepassing
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 a fraud detector can be made markedly better at two hard problems — camouflaged fraud and long-tailed rare fraud — by importing two mechanisms from hippocampal function: cross-view conflict detection and novelty-based attention. It builds a multi-view hypergraph over transactions (grouping temporally adjacent transactions per view key such as sender, receiver, location, type) and then couples a module that scores cross-view discrepancies with a module that reweights hypergraph message passing by local deviation from a neighborhood 'expectation.' The authors report that this combination consistently beats 15 state-of-the-art graph and non-graph baselines on six datasets, with average AUC, F1, and AP gains of 6.42%, 9.74%, and 39.14%, respectively. If these results hold, they establish a new state of the art for web finance fraud detection and a reusable design pattern: use local deviation as an explicit learning signal in graph neural networks for imbalanced, deceptive settings.

What carries the argument

The central object is a multi-view hypergraph built by temporal sliding windows: for each view key, transactions are sorted by timestamp and every four consecutive ones form a hyperedge. On this structure, three components carry the argument: (1) HCDP measures cross-view inconsistency per node via Jaccard distance on neighbor sets, Jensen–Shannon divergence on neighbor feature distributions, and entropy difference on neighbor label distributions; (2) CNHL computes a hyperedge center (CA3-like prediction), per-dimension variance (novelty), and a normalized novelty score that reweights the aggregated messages from neighbors before a sigmoid update; (3) MHF learns view-attention weights with mo

What would settle it

Vary the temporal window size w over a much wider range (e.g., 1, 8, 16, 32) on all six datasets and record the performance curve. If w=4 is not on a plateau or if the optimal w differs sharply across datasets, the fixed-window choice is a hidden lever in the reported SOTA performance. A second check: retrain the baselines that also operate on temporal graphs (e.g., GTAN, STAGN) with the same chronological split and label-ratio settings as HIMVH, and see whether HIMVH's AP lead persists when baselines are given per-dataset hyperparameter tuning.

Watch

Extended reading notes

Core claim

On six web-finance fraud datasets (two public, four private), HIMVH outperforms all 15 baselines on AUC, F1, and AP, with average gains of 6.42%, 9.74%, and 39.14% respectively. The authors attribute the gains to two mechanisms: HCDP, which computes structural, feature, and label discrepancies of each transaction across multiple views (sender, receiver, location, type) to expose imitation-based camouflage; and CNHL, which computes a hyperedge center as an expectation, measures per-dimension variance as novelty, and reweights messages so that rare tail fraud with atypical features gets amplified. Ablations show the novelty-aware hypergraph learning contributes the most, followed by the multi-

Load-bearing premise

The load-bearing premise is that grouping transactions into fixed-size temporal windows of four per view captures the behavioural context needed to define both cross-view discrepancies and neighborhood novelty; if fraud operates at longer or variable timescales, these hyperedges may misalign and the reported improvements may not transfer.

Editorial extensions

If this is right

  • HIMVH would give fraud teams a model that is better at ranking fraudulent transactions first, since the largest average gain is in average precision (39.14%), precisely the metric that matters when investigation capacity is limited.
  • The cross-view discrepancy module offers a concrete way to combat imitation-based fraud: by comparing a transaction's structural, feature, and label contexts across views whose attributes are hard to fake, the model can flag behavior that mimics normal patterns in one view but not in another.
  • The novelty-aware hypergraph learning provides a mechanism for long-tailed data: instead of treating all neighbors equally, it amplifies messages from transactions that deviate from their hyperedge's expected pattern, which the ablation shows is the single most important component.
  • Because the method only uses label entropy estimates from an auxiliary predictor rather than ground-truth neighbors, it is designed to work in semi-supervised settings, making it applicable to real deployments where labeled fraud is scarce.

Reading between the lines

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

  • The large gap between AP gains (39.14%) and AUC gains (6.42%) suggests the model's edge is concentrated in the ranking of tail fraud cases; a per-dataset decomposition of AP gains by fraud frequency would show whether the novelty module specifically rescues the rarest classes or also improves easy cases.
  • The fixed window size w=4 is an untested assumption across time scales; the natural extension is to make the window size per-view or per-category adaptive (or to use multi-scale windows), which would likely preserve or extend the gains on datasets with bursty or sparse transaction patterns.
  • The hippocampus framing is a design metaphor rather than a mechanistic model; the transferable insight is the variance-based reweighting rule, which could be dropped into any heterophilous graph model for other anomaly-detection tasks such as fake review detection, insurance claim fraud, or network intrusion, independent of the biological language.
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 paper proposes HIMVH, a hippocampus-inspired multi-view hypergraph learning model for web finance fraud detection. The model constructs per-view hypergraphs by temporally slicing transactions, computes cross-view identity/feature/label discrepancy features to counter camouflage, and uses a variance-based novelty reweighting message-passing scheme to handle long-tailed fraud. Experiments on six datasets (two public, four private) against 15 baselines are used to claim consistent state-of-the-art performance, with reported average improvements of 6.42% in AUC, 9.74% in F1, and 39.14% in AP. The paper also includes ablations, hyperparameter sensitivity, t-SNE visualizations, and SHAP analysis.

Significance. If the empirical claims are correct, HIMVH would be a meaningful advance for web finance fraud detection: the HCDP and CNHL modules are concrete, well-motivated mechanisms targeting camouflage and long-tailed distributions, and the benchmark is comparatively broad (six datasets, 15 baselines). The paper is strong in experimental breadth and includes ablations, sensitivity analysis, and interpretability. However, the headline quantitative claims contain an internal metric swap, several specific numbers in Section 5.2 are not supported by Table 2, all results are single-run with no uncertainty quantification, and the private datasets and lack of released code prevent independent verification. The hippocampus narrative is motivational rather than a formal derivation; this is not itself a flaw, but it means the contribution rests entirely on the empirical evaluation.

major comments (4)
  1. [Abstract; Conclusion; Section 5.2; Table 2] The reported improvement metrics are internally swapped. Recomputing the average relative improvement over the 15 baselines in Table 2 gives approximately AUC 6.6%, AP 10.0%, and F1 41.9%. The Abstract and Conclusion report '6.42% in AUC, 9.74% in F1 and 39.14% in AP'; the 9.74% figure corresponds to AP, and the 39.14% figure corresponds to F1. The same swap appears in Section 5.2 for the graph-learning baseline comparison (9.36% F1 vs 27.80% AP). The exact improvement formula is never stated, and as written the central quantitative contribution is mislabeled. Please correct all swapped metric names and provide the formula used.
  2. [Section 5.1, Table 2] All metrics in Table 2 are single-run numbers with no standard deviations, confidence intervals, or significance tests. Four of the six datasets are private and no code is released, so the claim that HIMVH 'consistently outperforms all baselines' cannot be independently checked. Some margins are small, e.g., AP on Private-1 is 92.91 vs 92.51 for the best baseline, and AUC on Private-2 is 99.17 vs 99.00 for UniGAD. Run-to-run variability could affect these conclusions. Please report mean plus/minus standard deviation over at least five seeds, perform significance tests against the best baseline, and release code (at least for the public datasets) with the final version.
  3. [Section 5.2, Table 2] The sentence 'Even on highly imbalanced datasets such as Private-3 and Private-4, HIMVH outperforms all baselines by at least 3.72% and 12.32% in AP, respectively' is not supported by Table 2. On Private-3, the largest AP margin over the best baseline is about 0.74 percentage points (86.34 vs GTAN 85.60); on Private-4, the margin over GTAN is about 5.2 percentage points (84.32 vs 79.08). These yield relative improvements of roughly 0.9% and 6.6%, not 3.72% and 12.32%. Please correct these numbers or explain the basis for them.
  4. [Section 4.2, Eqs. (7)-(8)] The label discrepancy term is defined inconsistently. Eq. (7) defines H_i^a from r_{i,label}, the empirical probability of ground-truth labels among neighbors, but the following sentence states that label entropy is computed using 'the predicted risk distribution from a lightweight auxiliary MLP instead of ground-truth neighbor labels.' These are different quantities. If ground-truth neighbor labels are used, the HCDP feature creates a label-leakage path in the semi-supervised setting; if predicted probabilities are used, the equation should be rewritten to reflect that. Please clarify and state how unlabeled nodes are handled in Eq. (7).
minor comments (6)
  1. [Section 4.3, Eq. (11)] The variable k in the variance formula is undefined; presumably it is the hyperedge size w. Please define it explicitly.
  2. [Section 4.3, Eq. (14)] The message aggregation uses alpha_j, described as a normalized novelty score derived from s_j, but the normalization formula is not given. Please provide it.
  3. [Algorithm 1, line 3] The loop variables are (a_j, a_k), but the equation inside the loop uses a_1, a_2. Please fix the notation for consistency.
  4. [Section 5.4, Figure 5] The sensitivity analysis does not state which dataset(s) it uses. If it is a single dataset, that limits the generalization claim; please specify and, ideally, include error bars or a second dataset.
  5. [Section 4.4, Eqs. (16)-(18)] The notation alpha_vs is introduced without definition, and the subscript vs is not explained. Please define all symbols in the fusion module.
  6. [Section 5.1] The paper says the F1 measure is 'macro average of F1 score,' which is unusual for fraud detection where the positive-class F1 is standard. Please clarify and, if macro-F1 is used, justify why.

Circularity Check

0 steps flagged · score 2.0 of 10

No material circularity: HIMVH's performance is empirical test-set measurement; the hippocampus framing is motivational analogy, and the few self-citations are contextual.

full rationale

The claimed derivation chain is not a derivation in the circularity-relevant sense. Sections 4.1-4.4 define a model: temporal-window hyperedge construction (Eqs. 1-3), cross-view discrepancy features (Eqs. 4-9), variance-based novelty reweighting (Eqs. 10-15), and view fusion (Eqs. 16-18). None of these equations are fitted to a target and then renamed as a prediction; they are fixed architectural operations. The central 'prediction' - HIMVH outperforming 15 baselines - is a direct empirical comparison on held-out test splits (Section 5), not a quantity forced by construction. The hippocampal material in Section 3 is explicitly presented as 'inspiration' and 'analogy' (e.g., 'serves as an analogy to the predictive pattern generated by CA3'), so it does not smuggle the result into the input. Self-citations [25], [43], and [53] are used only as related-work support for camouflage and graph-based fraud detection; the model equations and the evaluation do not depend on those papers, and no uniqueness theorem from prior work is invoked. I therefore find no circular step. Separately, the Abstract/Section 5.2 metric labels appear internally inconsistent with Table 2 (the 39.14% figure tracks F1 improvements and the ~9.7% figure tracks AP, not vice versa), but that is a reporting/consistency issue, not an input-output circularity.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The model's performance rests on several hand-set hyperparameters and domain assumptions about what constitutes fraud-relevant structure. The central claim is empirical, so these are the load-bearing choices.

free parameters (7)
  • temporal window size w = 4
    Controls the number of transactions per hyperedge; set to 4 without dataset-specific tuning (Eq. 2, §5.1).
  • novelty-sensitivity beta = 1
    Scales variance in novelty weight g_l; set to 1 (§5.1).
  • number of GNN layers = 3
    Set to 3 (§5.1); sensitivity analysis shows stable 3–5.
  • hidden dimension = 256
    Set to 256 (§5.1).
  • batch size = 512
    Set to 512 (§5.1).
  • learning rate = 0.001
    Adam optimizer lr=0.001 (§5.1).
  • number of views = 4
    Selected as 4 views (§5.1), though the view set A is not exhaustively specified.
assumptions (4)
  • domain assumption Transactions sorted by timestamp and grouped into fixed-size windows per view-specific key form hyperedges that encode fraud-relevant context (Eqs. 1–2).
    If temporal grouping at scale w does not align with fraud mechanisms, the model's inputs lose signal.
  • domain assumption Cross-view discrepancies (Jaccard distance, Jensen-Shannon divergence, label-entropy difference) are meaningful indicators of camouflaged fraud (Eqs. 4–8).
    No theoretical link between these metrics and fraud is derived; the paper relies on the intuition that fraudsters can't manipulate all views.
  • domain assumption Using label embeddings of labeled nodes in message passing is a valid semi-supervised strategy that does not leak test labels under the chronological split (Eq. 15, §4.3).
    The paper does not analyze potential train/test contamination through shared hyperedges or users.
  • domain assumption The 15 baselines were implemented and tuned to a fair, comparable standard (§5.1–5.2).
    No baseline hyperparameter details or code are provided, so the comparison rests on this assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bridging Cognitive Neuroscience and Graph Intelligence: Hippocampus-Inspired Multi-View Hypergraph Learning for Web Finance Fraud." pith.science (2026). https://pith.science/paper/A2QAUNP4

@misc{pith2026260111073,
  author       = {Pith},
  title        = {Pith review of: Bridging Cognitive Neuroscience and Graph Intelligence: Hippocampus-Inspired Multi-View Hypergraph Learning for Web Finance Fraud},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A2QAUNP4}},
  note         = {Machine review of arXiv:2601.11073}
}
read the original abstract

Online financial services constitute an essential component of contemporary web ecosystems, yet their openness introduces substantial exposure to fraud that harms vulnerable users and weakens trust in digital finance. Such threats have become a significant web harm that erodes societal fairness and affects the well-being of online communities. However, existing detection methods based on graph neural networks (GNNs) struggle with two persistent challenges: (1) long-tailed data distributions, which obscure rare but critical fraudulent cases, and (2) fraud camouflage, where malicious transactions mimic benign behaviors to evade detection. To fill these gaps, we propose HIMVH, a Hippocampus-Inspired Multi-View Hypergraph learning model for web finance fraud detection. Specifically, drawing inspiration from the scene conflict monitoring role of the hippocampus, we design a cross-view inconsistency perception module that captures subtle discrepancies and behavioral heterogeneity across multiple transaction views. This module enables the model to identify subtle cross-view conflicts for detecting online camouflaged fraudulent behaviors. Furthermore, inspired by the match-mismatch novelty detection mechanism of the CA1 region, we introduce a novelty-aware hypergraph learning module that measures feature deviations from neighborhood expectations and adaptively reweights messages, thereby enhancing sensitivity to online rare fraud patterns in the long-tailed settings. Extensive experiments on six web-based financial fraud datasets demonstrate that HIMVH achieves 6.42% improvement in AUC, 9.74% in F1 and 39.14% in AP on average over 15 SOTA models.

Figures

Figures reproduced from arXiv: 2601.11073 by the authors.

Figure 1
Figure 1. (a) Two attack strategies targeting the head and [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Two mechanisms in hippocampus. complex patterns [38, 43]. With the growing sophistication of on￾line fraudulent behaviors [28], graph neural networks (GNNs) have become a mainstream paradigm in web finance fraud detection due to their capability to capture and learn from the intricate relational structures embedded within web-based financial systems [14, 16]. Nevertheless, these web finance fraud detection methods b… view at source ↗
Figure 3
Figure 3. Overview of the proposed HIMVH model. However, traditional deep learning models still treat online trans￾actions as independent samples and overlook cross-entity relation￾ships. This limitation reduces their capability to capture coordi￾nated or camouflaged fraud. As a result, researchers increasingly adopt GNNs, which model structural dependencies and multi-hop interaction patterns to reveal sophisticated web-based… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: The ablation analysis on S-FFSD and Private-1. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Parameter sensitivity analysis. graph learning paradigm and mitigate over-smoothing encountered in deeper GNNs [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Embedding visualization of different models. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: SHAP analysis on Private-4. sparse and ambiguous. UniGAD unifies multi-level tasks through subgraph sampling and a shared fusion module. However, its sam￾pler prioritizes nodes with pronounced signals while discarding neighbors that appear normal yet convey essential r…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

56 extracted references · 3 canonical work pages

  1. [1]

    Laila Abd-Ellatif et al. 2025. ATAD-Net: An Adaptive Deep Learning Framework for Real-Time Financial Fraud Detection. Advances in Artificial Intelligence and Bridging Cognitive Neuroscience and Graph Intelligence: Hippocampus-Inspired Multi-View Hypergraph Learning for Web Finance Fraud Machine Learning, 2025; 5 (2): 225

  2. [2]

    Seyyede Zahra Aftabi, Ali Ahmadi, and Saeed Farzi. 2023. Fraud detection in financial statements using data mining and GAN models.Expert Systems with Applications227 (2023), 120144

  3. [3]

    Katie C Bittner, Christine Grienberger, Sachin P Vaidya, Aaron D Milstein, John J Macklin, Junghyup Suh, Susumu Tonegawa, and Jeffrey C Magee. 2015. Con- junctive input processing drives feature selectivity in hippocampal CA1 neurons. Nature neuroscience18, 8 (2015), 1133–1142

  4. [4]

    Nan Chen, Zemin Liu, Bryan Hooi, Bingsheng He, Rizal Fathony, Jun Hu, and Jia Chen. 2024. Consistency training with learnable data augmentation for graph anomaly detection with limited supervision. InThe Twelfth International Conference on Learning Representations

  5. [5]

    Dawei Cheng, Xiaoyang Wang, Ying Zhang, and Liqing Zhang. 2020. Graph neu- ral network for fraud detection via spatial-temporal attention.IEEE Transactions on Knowledge and Data Engineering34, 8 (2020), 3800–3813

  6. [6]

    Dawei Cheng, Sheng Xiang, Chencheng Shang, Yiyi Zhang, Fangzhou Yang, and Liqing Zhang. 2020. Spatio-temporal attention-based neural network for credit card fraud detection. InProceedings of the AAAI conference on artificial intelligence, Vol. 34. 362–369

  7. [7]

    Dawei Cheng, Yujia Ye, Sheng Xiang, Zhenwei Ma, Ying Zhang, and Changjun Jiang. 2023. Anti-Money Laundering by Group-Aware Deep Graph Learning. IEEE Transactions on Knowledge and Data Engineering35, 12 (2023), 12444–12457. doi:10.1109/TKDE.2023.3272396

  8. [8]

    Xiangyu Dong, Xingyi Zhang, Lei Chen, Mingxuan Yuan, and Sibo Wang. 2025. SpaceGNN: Multi-Space Graph Neural Network for Node Anomaly Detection with Extremely Limited Labels.arXiv preprint arXiv:2502.03201(2025)

Show all 56 references
  1. [9]

    Kang Fu, Dawei Cheng, Yi Tu, and Liqing Zhang. 2016. Credit card fraud de- tection using convolutional neural networks. InNeural Information Processing: 23rd International Conference, ICONIP 2016, Kyoto, Japan, October 16–21, 2016, Proceedings, Part III 23. Springer, 483–490

  2. [10]

    Yuan Gao, Xiang Wang, Xiangnan He, Zhenguang Liu, Huamin Feng, and Yong- dong Zhang. 2023. Addressing heterophily in graph anomaly detection: A per- spective of graph spectrum. InProceedings of the ACM Web Conference 2023. 1528–1538

  3. [11]

    Mingjian Guang, Zhong Li, Chungang Yan, Yuhua Xu, Junli Wang, Dawei Cheng, and Changjun Jiang. 2025. Multi-Temporal Partitioned Graph Attention Networks for Financial Fraud Detection.IEEE Transactions on Information Forensics and Security(2025)

  4. [12]

    Venus Haghighi, Behnaz Soltani, Nasrin Shabani, Jia Wu, Yang Zhang, Lina Yao, Quan Z Sheng, and Jian Yang. 2024. TROPICAL: Transformer-based Hyper- graph Learning for Camouflaged Fraudster Detection. In2024 IEEE International Conference on Data Mining (ICDM). IEEE, 121–130

  5. [13]

    Li Han, Longxun Wang, Ziyang Cheng, Bo Wang, Guang Yang, Dawei Cheng, and Xuemin Lin. 2025. Mitigating the Tail Effect in Fraud Detection by Com- munity Enhanced Multi-Relation Graph Neural Networks.IEEE Transactions on Knowledge and Data Engineering(2025)

  6. [14]

    Mengda Huang, Yang Liu, Xiang Ao, Kuan Li, Jianfeng Chi, Jinghua Feng, Hao Yang, and Qing He. 2022. Auc-oriented graph neural network for fraud detection. InProceedings of the ACM web conference 2022. 1311–1321

  7. [15]

    Joshua B Julian and Christian F Doeller. 2021. Remapping and realignment in the human hippocampal formation predict context-dependent spatial behavior. Nature neuroscience24, 6 (2021), 863–872

  8. [16]

    Abdullah Lakhan, Mazin Abed Mohammed, Dheyaa Ahmed Ibrahim, Seifedine Kadry, and Karrar Hameed Abdulkareem. 2022. ITS based on deep graph convolu- tional fraud detection network blockchain-enabled fog-cloud.IEEE Transactions on Intelligent Transportation Systems24, 8 (2022), 8399–8408

  9. [17]

    Xiangfeng Li, Shenghua Liu, Zifeng Li, Xiaotian Han, Chuan Shi, Bryan Hooi, He Huang, and Xueqi Cheng. 2020. FlowScope: Spotting Money Laundering Based on Graphs.Proceedings of the AAAI Conference on Artificial Intelligence34 (04 2020), 4731–4738. doi:10.1609/aaai.v34i04.5906

  10. [18]

    Yuqi Li, Guosheng Zang, Chunyao Song, and Xiaojie Yuan. 2025. A univer- sal adaptive algorithm for graph anomaly detection.Information Processing & Management62, 1 (2025), 103905

  11. [19]

    Yiqing Lin, Jianheng Tang, Chenyi Zi, H Vicky Zhao, Yuan Yao, and Jia Li

  12. [20]

    Junyi Liu, Dawei Cheng, and Changjun Jiang. 2023. Preventing Attacks in Interbank Credit Rating with Selective-aware Graph Neural Network. InIJCAI. 6085–6093

  13. [21]

    Yang Liu, Xiang Ao, Zidi Qin, Jianfeng Chi, Jinghua Feng, Hao Yang, and Qing He. 2021. Pick and choose: a GNN-based imbalanced learning approach for fraud detection. InProceedings of the web conference 2021. 3168–3177

  14. [22]

    Lundberg and Su-In Lee

    Scott M. Lundberg and Su-In Lee. 2017. A Unified Approach to Interpreting Model Predictions. InNeural Information Processing Systems. https://api.semanticscholar. org/CorpusID:21889700

  15. [23]

    Xuejiao Luo, Xiaohui Han, Wenbo Zuo, Xiaoming Wu, and Wenyin Liu. 2024. MLaD2: A Semi-Supervised Money Laundering Detection Framework Based on Decoupling Training.IEEE Transactions on Information Forensics and Security19 (2024), 4518–4533. doi:10.1109/TIFS.2024.3380262

  16. [24]

    Isaac Kofi Nti and Arjun Remadevi Somanathan. 2022. A scalable RF-XGBoost framework for financial fraud mitigation.IEEE Transactions on Computational Social Systems11, 2 (2022), 1556–1563

  17. [25]

    Rui Ou, Kun Zhu, Nana Zhang, MengChu Zhou, and Changjun Jiang. 2025. CamFD: Semi-Supervised Camouflage-Aware Fraud Detection Based on Dynamic Graphs.IEEE Transactions on Systems, Man, and Cybernetics: SystemsPP (01 2025), 1–15. doi:10.1109/TSMC.2025.3648571

  18. [26]

    Maxx Richard) Rahman, Lotfy Abdel Khaliq, Thomas Piper, Hans Geyer, Tristan Equey, Norbert Baume, Reid Aikin, and Wolfgang Maass. 2024. SACNN: self attention-based convolutional neural network for fraudulent behaviour detection in sports. InProceedings of the Thirty-Third Inte...

  19. [27]

    Y Sahin and Ekrem Duman. 2011. Detecting credit card fraud by decision trees and support vector machines. InProceedings of the International MultiConference of Engineers and Computer Scientists, Vol. 1. 1–6

  20. [28]

    Tayebi, Jian Pei, and Jiguo Cao

    Haolun Shi, Mohammad A. Tayebi, Jian Pei, and Jiguo Cao. 2023. Cost-Sensitive Learning for Medical Insurance Fraud Detection With Temporal Information . IEEE Transactions on Knowledge & Data Engineering35, 10 (Oct. 2023), 10451– 10463. doi:10.1109/TKDE.2023.3240431

  21. [29]

    Ajeet Singh, Anurag Jain, and Seblewongel Esseynew Biable. 2022. Financial fraud detection approach based on firefly optimization algorithm and support vector machine.Applied Computational Intelligence and Soft Computing2022, 1 (2022), 1468015

  22. [30]

    Emilija Strelcenia and Simant Prakoonwit. 2023. Improving Classification Perfor- mance in Credit Card Fraud Detection by Using New Data Augmentation.AI4 (01 2023), 172–198. doi:10.3390/ai4010008

  23. [31]

    Jianheng Tang, Fengrui Hua, Ziqi Gao, Peilin Zhao, and Jia Li. 2023. Gadbench: Revisiting and benchmarking supervised graph anomaly detection.Advances in Neural Information Processing Systems36 (2023), 29628–29653

  24. [32]

    Jianheng Tang, Jiajin Li, Ziqi Gao, and Jia Li. 2022. Rethinking graph neural networks for anomaly detection. InInternational Conference on Machine Learning. PMLR, 21076–21089

  25. [33]

    Yue Tian, Guanjun Liu, Jiacun Wang, and Mengchu Zhou. 2023. ASA-GNN: Adap- tive Sampling and Aggregation-Based Graph Neural Network for Transaction Fraud Detection.IEEE Transactions on Computational Social Systems(2023)

  26. [34]

    Rafaël Van Belle, Bart Baesens, and Jochen De Weerdt. 2023. CATCHM: A novel network-based credit card fraud detection method using node representation learning.Decision Support Systems164 (2023), 113866. doi:10.1016/j.dss.2022. 113866

  27. [35]

    Laurens Van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-SNE. Journal of machine learning research9, 11 (2008)

  28. [36]

    Chenxu Wang, Mengqin Wang, Xiaoguang Wang, Luyue Zhang, and Yi Long

  29. [37]

    Daixin Wang, Jianbin Lin, Peng Cui, Quanhui Jia, Zhen Wang, Yanming Fang, Quan Yu, Jun Zhou, Shuang Yang, and Yuan Qi. 2019. A semi-supervised graph at- tentive network for financial fraud detection. In2019 IEEE international conference on data mining (ICDM). IEEE, 598–607

  30. [38]

    Multi-Relational graph Representation learning for financial statement fraud detection.Big Data Mining and Analytics7, 3 (2024), 920–941

  31. [39]

    Xin Wang, Luo Xiangfeng, Xinzhi Wang, and Hang Yu. 2024. Homophilic and Heterophilic-Aware Sparse Graph Transformer for Financial Fraud Detection. In 2024 International Joint Conference on Neural Networks (IJCNN). 1–8. doi:10.1109/ IJCNN60899.2024.10650212

  32. [40]

    Shuhao Wang, Cancheng Liu, Xiang Gao, Hongtao Qu, and Wei Xu. 2017. Session- Based Fraud Detection in Online E-Commerce Transactions Using Recurrent Neural Networks. InMachine Learning and Knowledge Discovery in Databases, Yasemin Altun, Kamalika Das, Taneli Mielikäinen, Donat...

  33. [41]

    Jiasheng Wu, Xin Liu, Dawei Cheng, Yi Ouyang, Xian Wu, and Yefeng Zheng. 2024. Safeguarding fraud detection from attacks: A robust graph learning approach. InProceedings of the 33rd International Joint Conference on Artificial Intelligence. 7500–7508

  34. [42]

    Bin Wu, Xinyu Yao, Boyan Zhang, Kuo-Ming Chao, and Yinsheng Li. 2023. Split- GNN: Spectral Graph Neural Network for Fraud Detection against Heterophily. InProceedings of the 32nd ACM International Conference on Information and Knowledge Management. 2737–2746

  35. [43]

    Zhuojia Wu, Qi Zhang, Duoqian Miao, Xuerong Zhao, and Kaize Shi. 2025. Adapt- ing GNNs for Document Understanding: A Flexible Framework With Multiview Global Graphs.IEEE Transactions on Computational Social Systems12, 2 (2025), 608–621. doi:10.1109/TCSS.2024.3468890

  36. [44]

    Yanxi Wu, Liping Wang, Hongyu Li, and Jizhao Liu. 2025. A Deep Learning Method of Credit Card Fraud Detection Based on Continuous-Coupled Neural Networks.Mathematics13 (02 2025), 819. doi:10.3390/math13050819

  37. [45]

    Sheng Xiang, Guibin Zhang, Dawei Cheng, and Ying Zhang. 2025. Enhancing Attribute-Driven Fraud Detection With Risk-Aware Graph Representation.IEEE Transactions on Knowledge and Data Engineering(2025)

  38. [46]

    Huosong Xia, Wuyue An, and Zuopeng Justin Zhang. 2023. Credit risk models for financial fraud detection: A new outlier feature analysis method of xgboost with smote.Journal of Database Management (JDM)34, 1 (2023), 1–20. Rongkun Cui et al

  39. [47]

    Fei Xiao, Shaofeng Cai, Gang Chen, HV Jagadish, Beng Chin Ooi, and Meihui Zhang. 2024. VecAug: Unveiling Camouflaged Frauds with Cohort Augmentation for Enhanced Detection. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 6025–6036

  40. [48]

    Sheng Xiang, Mingzhi Zhu, Dawei Cheng, Enxia Li, Ruihui Zhao, Yi Ouyang, Ling Chen, and Yefeng Zheng. 2023. Semi-supervised credit card fraud detection via attribute-driven graph representation. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 37. 14557–14565

  41. [49]

    Biao Xu, Yao Wang, Xiuwu Liao, and Kaidong Wang. 2023. Efficient fraud detection using deep boosting decision trees.Decision Support Systems175 (2023), 114037

  42. [50]

    Yu Xie, Guanjun Liu, Chungang Yan, Changjun Jiang, and MengChu Zhou. 2023. Time-Aware Attention-Based Gated Network for Credit Card Fraud Detection by Extracting Transactional Behaviors.IEEE Transactions on Computational Social Systems10, 3 (2023), 1004–1016. doi:10.1109/TCSS....

  43. [51]

    Jinghui Zhang, Zhengjia Xu, Dingyang Lv, Zhan Shi, Dian Shen, Jiahui Jin, and Fang Dong. 2024. DiG-In-GNN: discriminative feature guided GNN-based fraud detector against inconsistencies in multi-relation fraud graph. InProceedings of the AAAI Conference on Artificial Intellige...

  44. [52]

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. 2018. How powerful are graph neural networks?arXiv preprint arXiv:1810.00826(2018)

  45. [53]

    Kun Zhu, Nana Zhang, Weiping Ding, and Changjun Jiang. 2024. An Adaptive Heterogeneous Credit Card Fraud Detection Model Based on Deep Reinforcement Training Subset Selection.IEEE Transactions on Artificial Intelligence5, 8 (2024), 4026–4041. doi:10.1109/TAI.2024.3359568

  46. [54]

    Yue Zhao and Maciej K Hryniewicki. 2018. Xgbod: improving supervised outlier detection with unsupervised representation learning. In2018 International Joint Conference on Neural Networks (IJCNN). IEEE, 1–8

  47. [56]

    Yao Zou and Dawei Cheng. 2025. Effective high-order graph representation learning for credit card fraud detection.arXiv preprint arXiv:2503.01556(2025)

  48. [2024]

    UniGAD: Unifying Multi-level Graph Anomaly Detection.arXiv preprint arXiv:2411.06427(2024)

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.