Pith. sign in

REVIEW 3 major objections 5 minor 32 references

Toward Auditable Fraud Detection: Combining Graph Features, Model Explanations, and Agentic Case Investigation

T0 review · 3 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read This paper claims that graph and anomaly features in a fraud pipeline improve detection only on hard cases and injected fraud rings, and that an LLM investigator producing clear rationales still makes more errors than the classifier it reli

desk verdict Honest PaySim study: after removing a simulator shortcut, graph/anomaly features don't help full-set AP, and an LLM agent underperforms its own classifier; the shortcut removal is plausible but under-validated. read the letter →

arxiv 2607.19266 v1 pith:GXHDTHG6 submitted 2026-07-21 cs.LG cs.AI

Rahil Sharma This is my paper
classification cs.LGcs.AI
keywords frauddetectionPaySimgraphfeaturesanomalyautoencoderLLMagentTreeSHAPescalationpolicy
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish when each layer of a fraud-detection system actually earns its place. On PaySim, after removing a simulator-specific balance shortcut, neither graph features nor an autoencoder signal improves average precision over the full test set, but both improve ranking among transactions the baseline scores uncertainly, and engineered graph features recover all injected multi-account ring transactions that the baseline often misses. The paper further claims that a bounded LLM investigation agent, given model explanations, graph context, and retrieved reference cases, underperforms a simple 0.5 threshold on the same classifier, converting six correct decisions into errors in eight disagreements while supplying a plausible written rationale each time. The underlying point: a component's value is conditional, and a coherent explanation is not evidence of a correct verdict.

What carries the argument

The load-bearing mechanism is the corrected baseline: the paper removes the two origin-balance fields (oldbalanceOrg and newbalanceOrig) plus the derived discrepancy, because a tree model could reconstruct from them a near-exact reconciliation with the transaction amount that the authors judge to be a PaySim simulator shortcut rather than a general fraud property. All comparisons—graph features, autoencoder signal, and the investigation agent—are then conditional on this corrected baseline, and the agent is bounded to tools derived from the same classifier's feature set, so it has no independent evidence with which to outperform the classifier by a wide margin.

What would settle it

Compute the origin-balance reconciliation statistic on a real, externally labelled mobile-money fraud dataset: if fraudulent transactions there also show near-exact reconciliation at high precision, the removed fields carry genuine signal; if they do not, the removal is justified.

Watch

Extended reading notes

Core claim

The central discovery, stated on the paper's own terms, is that a corrected tabular XGBoost baseline remains the strongest model on the full PaySim test set, while graph-derived structural features and an autoencoder reconstruction-error signal contribute only in narrower settings: the anomaly signal ranks fraud best within the fixed subset of intermediate-score cases, and the structural features recover all 39 injected ring transactions that the baseline detected inconsistently. The agentic layer, evaluated on a balanced 60-case sample, reaches 65.0% accuracy versus 71.7% for direct thresholding; six of its eight decision changes replaced correct classifier outputs with errors. The paper co

Load-bearing premise

The load-bearing premise is that the near-exact matching of origin balance fields and amount for fraudulent PaySim transactions is a simulator artifact rather than a genuine fraud indicator; if that matching also occurs in real fraudulent transactions, the corrected baseline is artificially handicapped and the comparisons that follow are unfair.

Editorial extensions

If this is right

  • On full transaction populations, adding graph or anomaly signals to a strong tabular baseline should not be expected to lift aggregate precision; gains appear in subsets the baseline finds uncertain.
  • Structural features are the part that pays off for recurring multi-account patterns: the graph-augmented model caught all injected ring transactions while the tabular baseline missed about a quarter, at a small cost in recall on the dataset's native fraud.
  • An LLM investigation agent that only re-packages classifier-derived evidence should not be granted override authority; here it degraded six of the eight decisions it changed, despite producing a coherent rationale each time.
  • Classifier-agent disagreement combined with a confident classifier score is a workable, transparent signal for escalating cases to human review; it flagged two agent errors in the sample without escalating a correct decision.

Reading between the lines

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

  • If the balance-shortcut assumption is wrong—if real fraudulent mobile-money transactions also show near-exact origin-balance reconciliation—then the corrected baseline is artificially handicapped and the full-set AP comparisons understate what a raw tabular model could do; this is testable on real data.
  • The agent's failure pattern suggests a general lesson for LLM-based investigators: when retrieval distances are computed separately per class and shown as evidence, an LLM may over-weight them against a calibrated classifier score; a calibration step or explicit uncertainty band around those distances might be a cheap fix.
  • The conditional-contribution finding implies that production systems may do better routing cases by detector specialty—graph features for relationship-based patterns, anomaly scores for outliers—rather than concatenating all signals into one model.
  • The disagreement-based escalation rule, which the authors note was derived on the same sample it was tested on, is a candidate for prospective validation: if it holds on new data, it could become a simple governance mechanism for LLM-agent deployment.
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

3 major / 5 minor

Summary. The paper proposes and evaluates a layered fraud-detection pipeline on the PaySim dataset: an XGBoost baseline, graph-derived structural and node2vec features, an autoencoder anomaly signal, TreeSHAP explanations, a bounded LLM investigation agent, and an escalation rule. Before model comparison the author removes two origin-balance fields, claiming they encode a simulator-specific shortcut. The main empirical findings are that, after this correction, the graph and anomaly additions do not improve full-test Average Precision, though they help on a baseline-defined intermediate-score subset; engineered structural features recover all injected synthetic fraud rings while the tabular baseline misses several; and the LLM agent makes fewer correct binary verdicts than directly thresholding the classifier it uses (39/60 vs 43/60), with six of eight disagreements converting correct decisions into errors. The paper concludes that each component contributes only under specific conditions and that a coherent agent rationale is not evidence of a better decision.

Significance. If the results hold, the paper provides a useful cautionary counterpoint to uncritical adoption of graph features, anomaly signals, and LLM investigation layers in fraud pipelines. Its strengths include explicit treatment of a data-quality shortcut, candid acknowledgment of transductive graph construction and of the conditional nature of the difficult-case subset, and a reproducible artifact with archived implementation details. The failure-pattern analysis of the agent (over-weighting reference distances) is a valuable qualitative contribution. However, the headline negative claims are built on a single random split, small samples, and an unvalidated preprocessing assumption, so the paper is best read as a well-hedged exploratory study rather than as definitive evidence about component value in operational fraud detection.

major comments (3)
  1. [§3.1, Appendix A] The full-test negative result depends on the claim that exact reconciliation between oldbalanceOrg, newbalanceOrig, and amount is a simulator artifact rather than a genuine fraud signal. The manuscript provides no quantitative evidence for this: no reconciliation rates for fraudulent vs legitimate transactions, no inspection of the PaySim generator source, and no external validation. If the reconciliation is actually a real property (e.g., fraudsters transferring the entire account balance), then removing the origin balance fields handicaps the baseline and biases every subsequent comparison. Please report the observed reconciliation statistics, cite or analyze the generator logic, and include a sensitivity analysis that retains the balance fields or tests the model variant where they are kept. This is load-bearing for the abstract's claim that 'neither the graph features nor the anomaly
  2. [§4.1, §5.1, Table 1] The central negative claim is derived from a single random stratified split. The full-test Average Precision differences are small (0.7677 vs 0.7611, 0.7662, 0.7617), and no confidence intervals, repeated splits, or significance tests are reported. With this design, the dataset supports only the statement that the additions did not improve AP in this particular split, not that they 'do not improve' AP in general. Please add bootstrap confidence intervals and/or repeated stratified and temporal splits, or explicitly weaken the central claim to an observed-sample statement. The same limitation applies to the difficult-subset comparison, where the positive class contains only 184 cases.
  3. [§5.3, Table 2] The agent comparison is based on 60 cases and a single run of a single LLM configuration. The difference between 39/60 and 43/60 is small, and on the eight disagreements a McNemar-style test would not reject the null hypothesis. The paper appropriately labels the sample as case-control and not deployment-representative, but the conclusion that the agent 'underperforms' direct thresholding is stronger than the evidence supports. Please provide repeated sampling, confidence intervals for the difference, and ideally multiple prompt/tool configurations; at minimum, frame the result as an observed pattern rather than a demonstrated disadvantage. The one-run limitation is acknowledged in §7, but it should also be reflected in the abstract and conclusion.
minor comments (5)
  1. [§5.3] The paragraph before Table 2 and the paragraph after the table repeat the same 2-out-of-8 vs 6-out-of-8 counts almost verbatim. Consider merging to avoid redundancy.
  2. [Figure 3] The two panels use different vertical scales; please state this explicitly in the caption and in the text, as the visual contrast can be misleading.
  3. [Appendix A] The helper function named 'train_test_split_by_step' does not perform a chronological split. Rename it or add a prominent warning to avoid confusion for readers who reuse the code.
  4. [§3.2] When describing zero vector embeddings for out-of-vocabulary accounts, clarify that this also affects the injected ring experiment, where all injected accounts are out-of-vocabulary. The text already notes this in §5.2, but an early pointer would help.
  5. [§7] The Limitations section is unusually thorough and helpful. One small gap: it does not revisit the balance-shortcut assumption from §3.1, which is the preprocessing step most vulnerable to external criticism.

Circularity Check

1 steps flagged · score 3.0 of 10

Core comparisons are independent; the only concrete circularity is the exploratory escalation rule, which is fit and evaluated on the same 60 cases.

  1. fitted input called prediction [Section 5.3 'Agent Decisions and Escalation'; acknowledged in Section 7 Limitations]
    "Based on this observed failure pattern, an exploratory escalation rule was applied after the agent evaluation. A case was passed to human review when the agent disagreed with the classifier decision and the classifier score was at least 0.65 or at most 0.35. ... Both escalated cases were agent errors, and no correct agent decisions were escalated in this sample. ... The escalation rule is exploratory and was developed after inspecting the same 60 cases on which it was evaluated."

    The escalation rule’s thresholds were chosen after inspecting the same 60-case sample on which its performance is then reported. The claim that the rule 'flagged two of these agent errors for human review without flagging any correct decision' is therefore a restatement of the post-hoc selection criterion, not an out-of-sample prediction. The paper explicitly concedes this: 'Its apparent ability to identify two agent errors without escalating a correct decision may therefore reflect the sample used to design it.' Because the rule is explicitly exploratory and secondary, this circularity does not force the paper’s central model-comparison or agent-vs-classifier conclusions.

full rationale

The main derivation chain is not circular. The full-test and difficult-subset model comparisons use a clean stratified train/test split with the same XGBoost configuration across variants, and the agent-vs-threshold comparison is a genuine out-of-sample evaluation on a balanced 60-case sample. The removal of the origin-balance fields in Section 3.1 is a data-preprocessing assumption about a PaySim-specific shortcut, not a definitional reduction of the stated conclusions; it is a validity risk rather than a circularity risk. The difficult subset is intentionally defined using the hard baseline’s own scores, and the paper candidly warns that this is conditional and not an unbiased subgroup comparison, but the AP values are computed, not derived tautologically from the subset definition. The one concrete circular element is the exploratory escalation rule, whose thresholds were fit and evaluated on the same 60 cases, producing a result that is in-sample by construction. Since the paper flags this limitation itself and does not use the escalation rule as a load-bearing claim for its central findings, the overall circularity score is modest.

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

No invented physical entities. The central claims inherit several domain assumptions: PaySim realism, the shortcut artifact, a single transductive split, and the evidential value of SHAP/reference distances. Free parameters are mostly standard ML settings, but the mid-score window and escalation thresholds are hand-selected or in-sample, and the class weight is data-derived.

free parameters (6)
  • XGBoost hyperparameters = 300 trees, max_depth=6, lr=0.1, seed=42
    Fixed by hand across all variants in Section 3.3; not tuned or cross-validated, but they define the classifier.
  • scale_pos_weight = ratio of legitimate to fraudulent training transactions
    Set to class ratio in Section 3.3; standard imbalance handling, depends on training labels.
  • difficult-case interval = [0.2, 0.8]
    Chosen by hand in Section 3.5 to define the agent evaluation set and the difficult subset; all conditional results depend on this window.
  • escalation rule thresholds = 0.35 and 0.65
    Set after inspecting the same 60 agent cases (Section 5.3); in-sample selection.
  • autoencoder hyperparameters = d->8->4->8->d, 20 epochs, lr=1e-3, batch=4096
    Chosen in Section 3.4 and Appendix A; affects the anomaly signal.
  • node2vec hyperparameters = dim=32, walk_length=20, walks=100, window=10
    Chosen in Appendix A; produces 64 embedding features per transaction.
assumptions (6)
  • domain assumption The near-perfect origin-balance reconciliation for fraud is a PaySim simulator artifact, not a real fraud signal.
    Invoked in Section 3.1 to justify dropping oldbalanceOrg and newbalanceOrig; if false, the corrected baseline is biased.
  • domain assumption PaySim fraud patterns and the synthetic injected rings are a valid testbed for graph/agent fraud components.
    Section 7 states PaySim does not reproduce the full range of operational fraud behaviours; the paper itself limits generalization.
  • domain assumption Transductive graph features computed on the full transaction graph before the split are an acceptable evaluation protocol.
    Sections 3.2 and 7 acknowledge that test-side relationships can influence graph representations; results are not inductive.
  • domain assumption One random stratified split with seed 42 suffices to compare model variants.
    Section 7 states no repeated splits or significance tests are reported; comparisons may be sensitive to the single split.
  • domain assumption TreeSHAP attributions and Euclidean reference distances provide meaningful evidence to the agent.
    Sections 3.5 and 4.2 treat these as evidence without validating their quality; the agent is observed to over-weight distances.
  • domain assumption The LLM (Llama 3.3 70B via Groq) returns tool results and verdicts as intended at temperature 0.1.
    Section 7 notes one run of one model/prompt; behaviour may vary across models and runs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Toward Auditable Fraud Detection: Combining Graph Features, Model Explanations, and Agentic Case Investigation." pith.science (2026). https://pith.science/paper/GXHDTHG6

@misc{pith2026260719266,
  author       = {Pith},
  title        = {Pith review of: Toward Auditable Fraud Detection: Combining Graph Features, Model Explanations, and Agentic Case Investigation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GXHDTHG6}},
  note         = {Machine review of arXiv:2607.19266}
}
read the original abstract

Fraud detection systems must scale with rising transaction volume while remaining explainable and reviewable. We study a layered pipeline on the PaySim dataset that combines a gradient-boosted classifier, graph-derived structural features, an autoencoder-based anomaly signal, TreeSHAP explanations, and a bounded LLM investigation agent applied to cases the classifier scores uncertainly. Before any model comparison, we identify and remove a simulator-specific balance shortcut that would otherwise inflate baseline performance. After this correction, neither the graph features nor the anomaly signal improves Average Precision on the full test set. Both, however, rank fraud better within the subset of cases receiving intermediate baseline scores. In a controlled experiment with injected multi-account fraud rings, engineered structural features recover all injected test transactions, while the tabular baseline misses roughly a quarter of them. The investigation agent underperforms direct thresholding of the classifier it relies on, reaching 65.0% accuracy against 71.7% on a balanced 60-case sample, despite having access to model explanations, graph context, and retrieved reference cases. Of the eight decisions the agent changed, six replaced correct classifier outputs with errors, and it produced a coherent written rationale in each case. An exploratory disagreement-based escalation rule flagged two of these agent errors for human review without flagging any correct decision. We conclude that each component of a layered fraud system contributes only under specific conditions, and that a plausible rationale from an investigation agent is not evidence of a better decision.

Figures

Figures reproduced from arXiv: 2607.19266 by the authors.

Figure 1
Figure 1. Overview of the proposed framework, from the operational tension motivating this work (Stage 0) through [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. System architecture of the fraud detection and investigation pipeline. Transaction data is represented through [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Average Precision across model variants on the complete test set and the baseline defined difficult subset. The [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 3 canonical work pages

  1. [1]

    Xgboost: A scal- able tree boosting system

    Tianqi Chen and Carlos Guestrin. Xgboost: A scal- able tree boosting system. InProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 785–794, 2016

    2016

  2. [2]

    Sage: An llm-driven self reflective agentic framework for fraud detection.arXiv preprint arXiv:2606.08146, 2026

    Yichen Chen, Siying Li, Yuhang Liang, Lijun Wang, and Renyang Liu. Sage: An llm-driven self reflective agentic framework for fraud detection.arXiv preprint arXiv:2606.08146, 2026

    arXiv 2026

  3. [3]

    Graph neural networks for financial fraud detec- tion: A review.Frontiers of Computer Science, 2025

    Dawei Cheng, Yao Zou, Sheng Xiang, and Changjun Jiang. Graph neural networks for financial fraud detec- tion: A review.Frontiers of Computer Science, 2025. doi: 10.1007/s11704-024-40474-y

    work page doi:10.1007/s11704-024-40474-y 2025

  4. [4]

    How paypal’s ai blocks$500 million in fraud per quarter, 2026

    Chief AI Officer. How paypal’s ai blocks$500 million in fraud per quarter, 2026. URL https://chiefaio fficer.com/how-paypals-ai-blocks-500-million -in-fraud-per-quarter/. Chief AI Officer

    2026

  5. [5]

    Ai fraud detection in banking 2026 guide,

    Emburse. Ai fraud detection in banking 2026 guide,

    2026

  6. [6]

    Experian’s new fraud forecast warns agentic ai, deepfake job candidates and cyber break-ins are top threats for 2026, 2026

    Experian plc. Experian’s new fraud forecast warns agentic ai, deepfake job candidates and cyber break-ins are top threats for 2026, 2026. URL https://www.ex perianplc.com/newsroom/press-releases/2026/e xperian-s-new-fraud-forecast-warns-agentic-a i--deepfake-job-can. Experian plc Newsroom

    2026

  7. [7]

    Friedman

    Jerome H. Friedman. Greedy function approximation: A gradient boosting machine.Annals of Statistics, pages 1189–1232, 2001. 11

    2001

  8. [8]

    node2vec: Scal- able feature learning for networks

    Aditya Grover and Jure Leskovec. node2vec: Scal- able feature learning for networks. InProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 855–864, 2016

    2016

Show all 32 references
  1. [9]

    Can llms find fraudsters? multi- level llm enhanced graph fraud detection

    Tairan Huang, Yili Wang, Qiutong Li, Changlong He, and Jianliang Gao. Can llms find fraudsters? multi- level llm enhanced graph fraud detection. InProceed- ings of the 33rd ACM International Conference on Multimedia (ACM MM ’25), pages 1530–1538, 2025. arXiv preprint arXiv:2507.11997

    2025

  2. [10]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations (ICLR), 2017

    2017

  3. [11]

    Se- fraud: Graph-based self-explainable fraud detection via interpretative mask learning

    Kaidi Li, Tianmeng Yang, Min Zhou, Jiahao Meng, Shendi Wang, Yihui Wu, Boshuai Tan, Hu Song, Lujia Pan, Fan Yu, Zhenli Sheng, and Yunhai Tong. Se- fraud: Graph-based self-explainable fraud detection via interpretative mask learning. InProceedings of the 30th ACM SIGKDD Confere...

    2024

  4. [12]

    Autonomous chain-of-thought distillation for graph-based fraud detection, 2026

    Yuan Li, Jun Hu, Bryan Hooi, Bingsheng He, and Cheng Chen. Autonomous chain-of-thought distillation for graph-based fraud detection, 2026. arXiv preprint arXiv:2601.22949

    2026

  5. [13]

    Graph-augmented large language model agents: Current progress and future prospects, 2026

    Yixin Liu, Guibin Zhang, Kun Wang, Shiyuan Li, and Shirui Pan. Graph-augmented large language model agents: Current progress and future prospects, 2026. arXiv preprint arXiv:2507.21407

    2026 arXiv

  6. [14]

    PaySim: A financial mobile money simulator for fraud detection

    Edgar Alonso Lopez-Rojas, Ahmad Elmir, and Stefan Axelsson. PaySim: A financial mobile money simulator for fraud detection. InProceedings of the 28th European Modeling and Simulation Symposium, Larnaca, Cyprus,

  7. [15]

    Lundberg and Su-In Lee

    Scott M. Lundberg and Su-In Lee. A unified approach to interpreting model predictions. InAdvances in Neu- ral Information Processing Systems (NeurIPS), 2017

    2017

  8. [16]

    Lundberg, Gabriel Erion, Hugh Chen, Alex DeGrave, Jordan M

    Scott M. Lundberg, Gabriel Erion, Hugh Chen, Alex DeGrave, Jordan M. Prutkin, Bala Nair, Ronit Katz, Jonathan Himmelfarb, Nisha Bansal, and Su-In Lee. From local explanations to global understanding with explainable AI for trees.Nature Machine Intelligence, 2(1):56–67, 2020. d...

    2020 doi

  9. [17]

    Sumit Misra, Harsh Thakur, Manas Ghosh, and Sami- ran K. Saha. An autoencoder based model for detecting fraudulent credit card transaction. InProcedia Com- puter Science, 2020. doi: 10.1016/j.procs.2020.03.219

    2020 doi

  10. [18]

    Agentic ai: Streamlining the future of ach fraud detection, 2026

    Nasdaq Verafin. Agentic ai: Streamlining the future of ach fraud detection, 2026. URL https://verafin.co m/2026/05/agentic-ai-streamlining-the-futur e-of-ach-fraud-detection/ . Nasdaq Verafin, 2026 Global Financial Crime Report

    2026

  11. [19]

    A comparison study of credit card fraud detection: Supervised versus unsupervised, 2019

    Xuetong Niu, Li Wang, and Xulei Yang. A comparison study of credit card fraud detection: Supervised versus unsupervised, 2019. arXiv preprint arXiv:1904.10604

    2019 arXiv

  12. [20]

    A label-free heterophily-guided approach for unsupervised graph fraud detection

    Junjun Pan, Yixin Liu, Xin Zheng, Yizhen Zheng, Alan Wee-Chung Liew, Fuyi Li, and Shirui Pan. A label-free heterophily-guided approach for unsupervised graph fraud detection. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 12443–12451, 2025. d...

    2025 doi

  13. [21]

    Correcting false alarms from unseen: Adapting graph anomaly detectors at test time

    Junjun Pan, Yixin Liu, Chuan Zhou, Fei Xiong, Alan Wee-Chung Liew, and Shirui Pan. Correcting false alarms from unseen: Adapting graph anomaly detectors at test time. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 24691–24699,

  14. [22]

    How paypal uses real-time graph database and graph analysis to fight fraud, 2021

    PayPal Engineering. How paypal uses real-time graph database and graph analysis to fight fraud, 2021. URL https://medium.com/paypal-tech/how-paypal-u ses-real-time-graph-database-and-graph-analy sis-to-fight-fraud-96a2b918619a

    2021

  15. [23]

    Kam, and Yee Ling Boo

    Tahereh Pourhabibi, Kok-Leong Ong, Booi H. Kam, and Yee Ling Boo. Fraud detection: A systematic literature review of graph-based anomaly detection approaches.Decision Support Systems, 133:113303,

  16. [24]

    doi: 10.1609/aaai.v40i29.39654

    doi

  17. [25]

    Stripe radar: fraud detection architecture and model choices

    Stripe. Stripe radar: fraud detection architecture and model choices. URL https://stripe.com/radar . Stripe

  18. [26]

    Explainability in graph neural networks: A taxonomic survey.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5782–5799, 2023

    Hao Yuan, Haiyang Yu, Shurui Gui, and Shuiwang Ji. Explainability in graph neural networks: A taxonomic survey.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5782–5799, 2023. doi: 10.1 109/TPAMI.2022.3204236

    2023

  19. [27]

    Agentic ai in payments in 2026: What’s real, what’s pilot and what’s still hype

    Nikita Zelezkins. Agentic ai in payments in 2026: What’s real, what’s pilot and what’s still hype. Finex- tra, February 2026. URL https://www.finextra.com /blogposting/30920/agentic-ai-in-payments-i n-2026-whats-real-whats-pilot-and-whats-sti ll-hype. Published 23 February 2026

    2026

  20. [28]

    The precision– recall plot is more informative than the ROC plot when evaluating binary classifiers on imbalanced datasets

    Takaya Saito and Marc Rehmsmeier. The precision– recall plot is more informative than the ROC plot when evaluating binary classifiers on imbalanced datasets. PLOS ONE, 10(3):e0118432, 2015. doi: 10.1371/jour nal.pone.0118432

    2015 doi

  21. [32]

    Let relations speak: An end-to-end llm-gnn soft prompt framework for fraud detection, 2026

    Zhixing Zuo, Huilin He, Jiasheng Wu, and Dawei Cheng. Let relations speak: An end-to-end llm-gnn soft prompt framework for fraud detection, 2026. arXiv preprint arXiv:2605.28524. 12 A Implementation Details T abular features.The corrected tabular representation contains the tr...

    2026 arXiv

  22. [2016]

    URL https://www.msc-les.org/proceedings /emss/2016/EMSS2016_249.pdf

    2016

  23. [2020]

    doi: 10.1016/j.dss.2020.113303

    2020

  24. [2026]

    URL https://www.emburse.com/resources/ ai-fraud-detection-in-banking. Emburse

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.