DGP: A Dual-Granularity Prompting Framework for Fraud Detection with Graph-Enhanced LLMs

Bingsheng He; Bryan Hooi; Cheng Chen; Jun Hu; Yuan Li

REVIEW 2 major objections 4 minor 2 cited by

DGP: A Dual-Granularity Prompting Framework for Fraud Detection with Graph-Enhanced LLMs

T0 review · 2 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Fraud detection on heterogeneous graphs can be handled by text-only LLM prompting when the target node keeps its full text and the neighbors are compressed into coarse summaries.

desk verdict A practical, well-executed LLM-for-fraud method whose main empirical claim hinges on split-construction details the paper never states; worth reviewing, but the authors need to clarify the protocol. read the letter →

arxiv 2507.21653 v1 pith:RMBITS63 submitted 2025-07-29 cs.LG cs.AI

Yuan Li , Jun Hu , Bryan Hooi , Bingsheng He , Cheng Chen This is my paper

classification cs.LGcs.AI

keywords frauddetectiongraph-enhancedLLMstext-onlypromptingdual-granularitymetapathsummarizationMarkovdiffusionkernelheterogeneousgraphAUPRC

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 is trying to establish that text-only LLM prompting can work well for fraud detection on heterogeneous graphs if the prompt is built at two granularities: the target node's own text stays intact and fine-grained, while all neighbor content is reduced to short coarse-grained summaries. The proposed framework, DGP, summarizes each node's text, prunes metapath neighborhoods with a Markov diffusion kernel, then summarizes the retained neighbor texts and mean-aggregates numerical features into per-metapath prompts. The empirical claim is that this asymmetry prevents the information overload that sinks full-text prompting and avoids the semantic loss of vector-encoding prompting, yielding up to a 6.8% AUPRC improvement over state-of-the-art baselines. A sympathetic reader should care because it suggests graph-enhanced LLMs can serve as standalone fraud classifiers with small token budgets, without graph-specific encoders.

What carries the argument

The central object is the dual-granularity prompt, which pairs the target node's original text with per-metapath coarse summaries. It is built by node-level textual summarization within a token budget, diffusion-based metapath trimming that keeps the top-M neighbors under the Markov diffusion kernel distance, metapath-level textual summarization that fuses the selected node summaries into one short text, and numerical summarization that mean-aggregates neighbor attributes. The machinery works because it controls two quantities in the attention-dilution analysis, the number of neighbor tokens m and the neighborhood size nK, keeping the fraud-related token fraction high.

What would settle it

Inspect the four datasets' splits for overlap, build a strict disjoint split where no user, product, or shop appears in more than one split, rerun DGP and the best baselines, and compare AUPRC; if the gap substantially shrinks, the leakage rather than the dual-granularity design is the driver.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that neighbor information in fraud graphs can be compressed almost arbitrarily without losing its predictive value, as long as the target node keeps its verbose self-description. DGP's prompt for a node consists of the node's raw text plus, for each metapath, a short textual summary of the selected neighbors and a mean vector of their numerical features. The framework selects which neighbors to summarize using Markov-diffusion-kernel distances, keeping only the top-M structurally and semantically closest nodes per metapath. Across YelpReviews, AmazonVideo, and two proprietary industry datasets, DGP achieves the best AUPRC on all four datasets, with gains of up to 6.8% over the strongest baselines.

Load-bearing premise

The claim depends on the labeled train, validation, and test splits being leak-free: the paper never states whether splits are random, temporal, or disjoint on users and products, so if the same reviewer or product appears in both training and test sets, the reported AUPRC gains could come from memorization rather than from dual-granularity prompting.

Editorial extensions

If this is right

Text-only graph-to-prompt fraud detection can scale to dense heterogeneous graphs: prompt length scales with the number of metapaths and hop count, not with neighborhood degree.
Neighbor text can be compressed to about 10 tokens per metapath without hurting accuracy; verbose neighbor descriptions are unnecessary and can dilute signal.
LLM-based fraud classifiers can outperform specialized GNNs on review and industry graphs when the prompt gives the target node fine detail and the neighborhood coarse context.
The attention-dilution analysis predicts that methods with full multi-hop neighbor concatenation will degrade faster on high-degree graphs, making DGP's advantage grow with graph density.
The dual-granularity design separates cleanly into components—text summarization, neighbor trimming, and numerical aggregation—each removable, and each removal is shown to lower performance.

Reading between the lines

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

If the attention-dilution mechanism is the real driver, DGP's margin over full-text prompting should grow monotonically with average out-degree; a synthetic experiment that varies degree while holding text content fixed would test that prediction directly.
The same dual-granularity recipe, target details verbatim with context summarized, could transfer to other text-attributed classification tasks such as spam, misinformation, or toxic-content detection, where the target's own text carries most of the signal.
Mean aggregation of numerical neighbor features assumes each neighbor contributes equally; replacing it with order statistics or a fraud-ratio estimate might catch schemes where one anomalous neighbor is buried among many benign ones.
The finding that task-agnostic summarization beats fraud-aware summarization suggests the LLM's generic summarization already preserves the cues needed for downstream classification; comparing which n-grams survive in the 10-token summaries of fraudulent versus benign neighborhoods could reveal what those cues are.

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

DGP is a practical, well-executed method paper. The dual-granularity idea is not conceptually deep, but the implementation is careful and the empirical wins are consistent across four datasets. The thing to watch before you trust the numbers: Section 5.1 never says how the train/validation/test splits are built. That is a real hole for an empirical paper, because the prompts include metapath neighborhoods (same user, same product) and a random node-level split could put the same user or product on both sides of the boundary. The authors note the split sizes don't cover all nodes, but they don't say whether splits are random, temporal, or disjoint on users/products. If there's identity leakage, the LLM could memorize user/product cues and inflate the text-based gains. This is the main weakness; it's fixable by stating the protocol and ideally releasing the split code.

What the paper does well: the fine-grained target + coarse-grained neighbor design directly addresses the information overload problem in text-only prompting, and the token-budget analysis is honest. The ablations show each component matters, and the observation that task-aware summarization hurts is a useful negative result. Related work is fair. The industrial datasets from ByteDance are a nice addition, though they also mean parts of the evaluation can't be independently reproduced.

The minor issue: Section 4.6 claims the cumulative attention on fraud-related tokens is 'exactly r'. That is not true; attention weights depend on content, not just token counts. The qualitative intuition about dilution is fine, but the 'exactly' should go. Also, no code or data release, which weakens the paper but is a norm problem, not a fatal one.

Overall, this is a legitimate incremental contribution, not a paradigm shift. The central claim rests on the unstated split protocol, so I can't call it fully sound as written, but the experiments look plausible and the method is sensible. I would send this to peer review and ask for the split details and a correction to the attention statement. If the splits turn out to be clean, this is a publishable paper.

Referee Report

2 major / 4 minor

Summary. The paper proposes Dual Granularity Prompting (DGP), a graph-to-prompt framework that uses LLMs as standalone classifiers for fraud detection on heterogeneous graphs. DGP preserves the target node's raw text at fine granularity while summarizing neighbor information into coarse-grained text and numerical prompts. The method combines bi-level textual summarization, Markov-diffusion-kernel-based metapath trimming, and statistical aggregation of numerical features. The authors evaluate DGP on two public review graphs and two industrial datasets, reporting consistent improvements over GNN, LLM, and graph-enhanced-LLM baselines, with AUPRC gains up to 6.8 percentage points, and they include ablations, token-budget analysis, and complexity estimates.

Significance. If the empirical results hold, DGP is a useful contribution to graph-enhanced LLMs for fraud detection: it demonstrates that a text-only prompting approach can handle heterogeneous fraud graphs with dense textual neighborhoods, and it provides a practical token-budget control mechanism. The paper's strengths include evaluation across four datasets including two industrial graphs, comparison with a wide range of baselines implemented from official code, ablation studies for each component, and explicit complexity analysis. The central claim is empirical and is supported by a consistent pattern in Table 2, but the unspecified train/validation/test split protocol leaves a load-bearing gap that must be addressed before the results can be trusted as measuring generalization to unseen fraud.

major comments (2)

[§5.1 (Datasets, Table 1)] Section 5.1 specifies only the sizes of the train/validation/test splits and states that the sum of split sizes can be smaller than the total number of nodes; it never states whether the splits are random, temporal, or disjoint on users/products, nor whether unlabeled nodes are used when constructing metapath neighborhoods for training and test nodes. On YelpReviews and AmazonVideo, a random review-level split can place reviews from the same user or on the same product in both training and test sets, and because DGP explicitly includes metapath neighborhoods (R-U-R, R-P-R, R-S-R) in the prompt, the model could exploit identity-level cues rather than learn to detect fraud among genuinely unseen users/products. This is load-bearing for the headline AUPRC gains in Table 2. Please specify the exact split-construction protocol, including whether unlabeled nodes are excluded from prompts, and if the current splits are not disjoint on users/products or temporal, re-run the experiments with leakage-free splits and report the resulting numbers.
[§4.6 (Eqs. 15–16)] The derivation claims that the cumulative softmax attention assigned to fraud-related tokens is 'exactly r', where r is the token fraction defined in Eq. (15). This is not valid: softmax attention weights α_i in Eq. (16) depend on the content-dependent query and key vectors, not only on token counts, and the equality holds only under a uniform-attention assumption that is neither stated nor justified. The subsequent discussion of attention dispersion is relevant, but the exact equality is false. This issue does not by itself invalidate the empirical comparison, but it makes the theoretical motivation in Section 4.6 inaccurate; please revise this section to present the token-fraction argument as a heuristic or explicitly state and justify the uniform-attention condition.

minor comments (4)

[§5.1 (Parameter Settings)] The sentence 'we tune the the number of hops' contains a duplicated article 'the'; it should read 'we tune the number of hops'.
[Figure 4] The label 'AURPC' in the ablation figure should be 'AUPRC' to match the metric name used throughout the paper.
[§4.5 (Complexity Analysis)] The prompt-length formula 'L + R(K+1)-R/(R-1) M B tokens' is ambiguous; adding parentheses, such as L + ((R^{K+1}-R)/(R-1)) · M · B, would clarify that the geometric factor multiplies both M and B.
[§4.2 (Node-level Summarization)] The observation that task-specific summarization prompts can underperform is stated without a reference or a detailed explanation; since Table 3 provides supporting evidence, consider moving or elaborating this discussion in the experiments section.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DGP is a concrete graph-to-prompt construction evaluated on held-out test nodes, and the flagged theoretical claim is a correctness issue rather than a circular reduction.

full rationale

The paper's derivation chain is self-contained and not circular. DGP defines its pipeline procedurally: node-level summarization (Eq. 5), diffusion-based metapath trimming (Eqs. 6-9), metapath summarization (Eq. 10), numerical aggregation (Eq. 11), prompt construction (Eq. 12), fine-tuning (Eq. 13), and inference calibration (Eq. 14). None of these equations fits a parameter to the evaluation target and then reports that fit as a prediction; hyperparameters are tuned on validation performance, and all reported metrics are computed on test splits. The citation to the authors' prior RpHGNN work appears only as motivational inspiration for retaining fine-grained target features while abstracting neighbors, and it is not load-bearing for any derived result. The same LLM family is used for summarization and classification, but this is a methodological choice, not a circular reduction: the summarization step does not consume labels or test information, and the classification model is trained only on labeled training prompts. The paper's Section 4.6 claim that cumulative softmax attention assigned to fraud-related tokens is 'exactly r' is mathematically unsupported because softmax attention weights are content-dependent, but this is a correctness flaw in a heuristic motivation, not a circularity in the empirical claim. Similarly, the unspecified train/test split construction is an external-validity and potential leakage concern, not a circular reasoning step. Therefore no circularity is found, and the score is 0.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The method relies on several tuned hyperparameters (B_node, B_meta, K, M, LoRA settings) and on untested assumptions about attention uniformity, semantic similarity via diffusion, and summarization fidelity. No new physical or conceptual entities are introduced.

free parameters (5)

B_node (node-level summarization token budget) = tuned per dataset in {10, 20, 40, 80}; final values not reported
Controls how much each node's text is compressed before metapath aggregation; directly affects prompt content and length.
B_meta (metapath-level summarization token budget) = tuned per dataset in {10, 20, 40, 80}
Controls the length of each metapath summary in the final prompt.
K (number of hops for Markov diffusion kernel) = tuned per dataset in {1, 2, 3}
Sets the averaging window for the diffusion operator in metapath trimming.
M (neighbor truncation size) = tuned per dataset in {2, 4, 8, 16}
Limits the number of retained neighbors per metapath after trimming.
LoRA rank, LoRA dropout, learning rate = grid search ranges given; final values not reported
Standard fine-tuning hyperparameters for the LLM backbone.

assumptions (3)

ad hoc to paper Softmax attention weights in the LLM are approximately uniform, so that the cumulative attention on fraud-related tokens equals their token fraction r (Eq. 15-16).
This is used to argue that long neighbor text dilutes fraud signals; it is not stated as an assumption and is generally false for trained transformers.
domain assumption The Markov diffusion kernel with raw node features produces embeddings that capture semantic similarity relevant to fraud, and the top-M closest neighbors under this metric are the most informative.
The method assumes that diffusion distance over features is a good proxy for fraud-relevant neighbor relevance; no evidence is provided for this specific task.
domain assumption Task-agnostic LLM summarization with a generic instruction preserves the fraud-relevant content of node texts within a 10-token budget.
The paper shows empirically that task-aware summaries hurt, but does not show that the task-agnostic summaries retain the specific cues needed for fraud classification.

how reviews work

0 comments

Cite this review

Pith. "Pith review of DGP: A Dual-Granularity Prompting Framework for Fraud Detection with Graph-Enhanced LLMs." pith.science (2026). https://pith.science/paper/RMBITS63

@misc{pith2026250721653,
  author       = {Pith},
  title        = {Pith review of: DGP: A Dual-Granularity Prompting Framework for Fraud Detection with Graph-Enhanced LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RMBITS63}},
  note         = {Machine review of arXiv:2507.21653}
}

read the original abstract

Real-world fraud detection applications benefit from graph learning techniques that jointly exploit node features, often rich in textual data, and graph structural information. Recently, Graph-Enhanced LLMs emerge as a promising graph learning approach that converts graph information into prompts, exploiting LLMs' ability to reason over both textual and structural information. Among them, text-only prompting, which converts graph information to prompts consisting solely of text tokens, offers a solution that relies only on LLM tuning without requiring additional graph-specific encoders. However, text-only prompting struggles on heterogeneous fraud-detection graphs: multi-hop relations expand exponentially with each additional hop, leading to rapidly growing neighborhoods associated with dense textual information. These neighborhoods may overwhelm the model with long, irrelevant content in the prompt and suppress key signals from the target node, thereby degrading performance. To address this challenge, we propose Dual Granularity Prompting (DGP), which mitigates information overload by preserving fine-grained textual details for the target node while summarizing neighbor information into coarse-grained text prompts. DGP introduces tailored summarization strategies for different data modalities, bi-level semantic abstraction for textual fields and statistical aggregation for numerical features, enabling effective compression of verbose neighbor content into concise, informative prompts. Experiments across public and industrial datasets demonstrate that DGP operates within a manageable token budget while improving fraud detection performance by up to 6.8% (AUPRC) over state-of-the-art methods, showing the potential of Graph-Enhanced LLMs for fraud detection.

Figures

Figures reproduced from arXiv: 2507.21653 by the authors.

**Figure 2.** Fraud detection performance (↑) vs. token usage per prompt (↓) across different methods and datasets. Our proposed method, DGP, achieves top performance with moderate token consumption, demonstrating a notable balance between token usage and performance. which incorporate LLM-encoded features and rely heavily on the classification capabilities of GNNs. In this work, we focus on leveraging graph-enhanced LLMs as st… view at source ↗

**Figure 3.** Overview of the proposed DGP framework. Another line of work, LLM-enhanced GNNs, integrates LLM-encoded features into GNNs to improve node representation. For example, TAPE (He et al. 2024) uses LLMgenerated explanations as auxiliary features for downstream GNNs, and FLAG (Yang et al. 2025) leverages discriminative text extraction to address neighborhood camouflage in fraud detection. While these approaches rely … view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Ablation study on DGP components. “Path” de [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: presents fraud detection metrics across a range of budgets B ∈ {5, 10, 20, 40, 80} for the YelpReviews and AmazonVideo datasets. For simplicity, we assume a unified budget, i.e., Bnode = Bmeta. We observe that very short summaries (5 tokens) provide insufficient conte…

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Operational Evidence Gaps for LLMs in Fraud Detection and Trust-and-Safety Workflows
cs.CR 2026-07 conditional novelty 5.0 of 10

A survey of 49 LLM fraud and trust-and-safety papers finds that fraud work reports almost no per-decision latency, cost, or calibration evidence, while moderation work reports more.
SHERLOCK: Towards Dynamic Knowledge Adaptation in LLM-enhanced E-commerce Risk Management
cs.IR 2025-10 conditional novelty 5.0 of 10

A deployed LLM risk-investigation system that augments retrieval and reflection with a domain knowledge base achieves 0.92 factual alignment and 82% expert acceptance at JD.com.

Reference graph

Works this paper leans on

42 extracted references · 23 canonical work pages · cited by 2 Pith papers

[1]

, " * write output.state after.block = add.period write newline

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...
[2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...
[3]

Akoglu, L.; Tong, H.; and Koutra, D. 2015. Graph based anomaly detection and description: a survey. Data mining and knowledge discovery, 29(3): 626--688

2015
[4]

Barbero, F.; Banino, A.; Kapturowski, S.; Kumaran, D.; Madeira Ara \'u jo, J.; Vitvitskyi, O.; Pascanu, R.; and Veli c kovi \'c , P. 2024. Transformers need glasses! information over-squashing in language tasks. Advances in Neural Information Processing Systems, 37: 98111--98142

work page 2024
[5]

Chen, N.; Liu, Z.; Hooi, B.; He, B.; Fathony, R.; Hu, J.; and Chen, J. 2024. Consistency training with learnable data augmentation for graph anomaly detection with limited supervision. In The twelfth international conference on learning representations

work page 2024
[6]

Dou, Y.; Liu, Z.; Sun, L.; Deng, Y.; Peng, H.; and Yu, P. S. 2020. Enhancing graph neural network-based fraud detectors against camouflaged fraudsters. In Proceedings of the 29th ACM international conference on information & knowledge management, 315--324

work page 2020
[7]

Duan, M.; He, D.; Zheng, T.; Jia, L.; Song, M.; Wang, X.; and Feng, Z. 2025. Global Attribute-Association Pattern Aggregation for Graph Fraud Detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, 11616--11624

work page 2025
[8]

Duan, M.; Zheng, T.; Gao, Y.; Wang, G.; Feng, Z.; and Wang, X. 2024. Dga-gnn: Dynamic grouping aggregation gnn for fraud detection. In Proceedings of the AAAI conference on artificial intelligence, volume 38, 11820--11828

work page 2024

Show all 42 references

[9]

Fatemi, B.; Halcrow, J.; and Perozzi, B. 2023. Talk like a graph: Encoding graphs for large language models. arXiv preprint arXiv:2310.04560

2023 arXiv
[10]

Fouss, F.; Francoisse, K.; Yen, L.; Pirotte, A.; and Saerens, M. 2012. An experimental investigation of kernels on graphs for collaborative recommendation and semisupervised classification. Neural networks, 31: 53--72

2012
[11]

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

2017
[12]

He, X.; Bresson, X.; Laurent, T.; Perold, A.; LeCun, Y.; and Hooi, B. 2024. Harnessing Explanations: LLM -to- LM Interpreter for Enhanced Text-Attributed Graph Representation Learning. In The Twelfth International Conference on Learning Representations

2024
[13]

A.; Beutel, A.; Shah, N.; and Faloutsos, C

Hooi, B.; Shin, K.; Song, H. A.; Beutel, A.; Shah, N.; and Faloutsos, C. 2017. Graph-based fraud detection in the face of camouflage. ACM Transactions on Knowledge Discovery from Data (TKDD), 11(4): 1--26

2017
[14]

J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; Chen, W.; et al

Hu, E. J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; Chen, W.; et al. 2022. Lora: Low-rank adaptation of large language models. ICLR, 1(2): 3

2022
[15]

Hu, J.; Hooi, B.; and He, B. 2024. Efficient heterogeneous graph learning via random projection. IEEE Transactions on Knowledge and Data Engineering

2024
[16]

Hu, Z.; Dong, Y.; Wang, K.; and Sun, Y. 2020. Heterogeneous graph transformer. In Proceedings of the web conference 2020, 2704--2710

2020
[17]

N.; and Welling, M

Kipf, T. N.; and Welling, M. 2017. Semi-Supervised Classification with Graph Convolutional Networks. In International Conference on Learning Representations

2017
[18]

Li, K.; Yang, T.; Zhou, M.; Meng, J.; Wang, S.; Wu, Y.; Tan, B.; Song, H.; Pan, L.; Yu, F.; et al. 2024. Sefraud: Graph-based self-explainable fraud detection via interpretative mask learning. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Min...

2024
[19]

Li, S.; Yang, J.; Liang, G.; Li, T.; and Zhao, K. 2022. SybilFlyover: Heterogeneous graph-based fake account detection model on social networks. Knowledge-Based Systems, 258: 110038

2022
[20]

Li, Y.; Li, Z.; Wang, P.; Li, J.; Sun, X.; Cheng, H.; and Yu, J. X. 2023. A survey of graph meets large language model: Progress and future directions. arXiv preprint arXiv:2311.12399

2023 arXiv
[21]

F.; Lin, K.; Hewitt, J.; Paranjape, A.; Bevilacqua, M.; Petroni, F.; and Liang, P

Liu, N. F.; Lin, K.; Hewitt, J.; Paranjape, A.; Bevilacqua, M.; Petroni, F.; and Liang, P. 2024 a . Lost in the Middle: How Language Models Use Long Contexts. Transactions of the Association for Computational Linguistics, 12: 157--173

2024
[22]

Liu, S.; Yao, D.; Fang, L.; Li, Z.; Li, W.; Feng, K.; Ji, X.; and Bi, J. 2024 b . Anomalyllm: Few-shot anomaly edge detection for dynamic graphs using large language models. In 2024 IEEE International Conference on Data Mining (ICDM), 785--790. IEEE

2024
[23]

Loshchilov, I.; and Hutter, F. 2019. Decoupled Weight Decay Regularization. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019 . OpenReview.net

2019
[24]

J.; and Leskovec, J

McAuley, J. J.; and Leskovec, J. 2013. From amateurs to connoisseurs: modeling the evolution of user expertise through online reviews. In Proceedings of the 22nd international conference on World Wide Web, 897--908

2013
[25]

Pedregosa, F.; Varoquaux, G.; Gramfort, A.; Michel, V.; Thirion, B.; Grisel, O.; Blondel, M.; Prettenhofer, P.; Weiss, R.; Dubourg, V.; Vanderplas, J.; Passos, A.; Cournapeau, D.; Brucher, M.; Perrot, M.; and Duchesnay, E. 2011. Scikit-learn: Machine Learning in P ython. Journ...

2011
[26]

Rayana, S.; and Akoglu, L. 2015. Collective opinion spam detection: Bridging review networks and metadata. In Proceedings of the 21th acm sigkdd international conference on knowledge discovery and data mining, 985--994

2015
[27]

Rosenblatt, F. 1958. The perceptron: a probabilistic model for information storage and organization in the brain. Psychological review, 65(6): 386

1958
[28]

Sharma, V.; Kumar, R.; Cheng, W.-H.; Atiquzzaman, M.; Srinivasan, K.; and Zomaya, A. Y. 2018. NHAD: Neuro-fuzzy based horizontal anomaly detection in online social networks. IEEE Transactions on Knowledge and Data Engineering, 30(11): 2171--2184

2018
[29]

S.; and Wu, T

Sun, Y.; Han, J.; Yan, X.; Yu, P. S.; and Wu, T. 2011. Pathsim: Meta path-based top-k similarity search in heterogeneous information networks. Proceedings of the VLDB Endowment, 4(11): 992--1003

2011
[30]

Tang, J.; Yang, Y.; Wei, W.; Shi, L.; Su, L.; Cheng, S.; Yin, D.; and Huang, C. 2024 a . Graphgpt: Graph instruction tuning for large language models. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, 491--500

2024
[31]

Tang, J.; Yang, Y.; Wei, W.; Shi, L.; Xia, L.; Yin, D.; and Huang, C. 2024 b . Higpt: Heterogeneous graph language model. In Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining, 2842--2853

2024
[32]

Team, Q. 2025. Qwen3 Technical Report. arXiv:2505.09388

2025 arXiv
[33]

N.; Kaiser, .; and Polosukhin, I

Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, .; and Polosukhin, I. 2017. Attention is all you need. Advances in neural information processing systems, 30

2017
[34]

Vasylenko, P.; Treviso, M.; and Martins, A. F. 2025. Long-Context Generalization with Sparse Attention. arXiv preprint arXiv:2506.16640

2025
[35]

Veličković, P.; Cucurull, G.; Casanova, A.; Romero, A.; Liò, P.; and Bengio, Y. 2018. Graph Attention Networks. In International Conference on Learning Representations

2018
[36]

Wang, H.; Feng, S.; He, T.; Tan, Z.; Han, X.; and Tsvetkov, Y. 2023. Can language models solve graph problems in natural language? Advances in Neural Information Processing Systems, 36: 30840--30861

2023
[37]

Xu, F.; Wang, N.; Wu, H.; Wen, X.; Zhao, X.; and Wan, H. 2024. Revisiting graph-based fraud detection in sight of heterophily and spectrum. In Proceedings of the AAAI conference on artificial intelligence, volume 38, 9214--9222

2024
[38]

Yang, C.; Liu, H.; Wang, D.; Zhang, Z.; Yang, C.; and Shi, C. 2025. FLAG: Fraud Detection with LLM-enhanced Graph Neural Network. In Proceedings of the 31st ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD '25)

2025
[39]

Ye, R.; Zhang, C.; Wang, R.; Xu, S.; and Zhang, Y. 2024. Language is All a Graph Needs. EACL

2024
[40]

Yu, H.; Liu, Z.; and Luo, X. 2024. Barely supervised learning for graph-based fraud detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 16548--16557

2024
[41]

Zhu, X.; Xue, H.; Zhao, Z.; Xu, W.; Huang, J.; Guo, M.; Wang, Q.; Zhou, K.; and Zhang, Y. 2025. Llm as gnn: Graph vocabulary learning for text-attributed graph foundation models. arXiv preprint arXiv:2503.03313

2025
[42]

Zhuo, W.; Liu, Z.; Hooi, B.; He, B.; Tan, G.; Fathony, R.; and Chen, J. 2024. Partitioning message passing for graph fraud detection. arXiv preprint arXiv:2412.00020

2024 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

, " * write output.state after.block = add.period write newline

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

[2] [2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

[3] [3]

Akoglu, L.; Tong, H.; and Koutra, D. 2015. Graph based anomaly detection and description: a survey. Data mining and knowledge discovery, 29(3): 626--688

2015

[4] [4]

Barbero, F.; Banino, A.; Kapturowski, S.; Kumaran, D.; Madeira Ara \'u jo, J.; Vitvitskyi, O.; Pascanu, R.; and Veli c kovi \'c , P. 2024. Transformers need glasses! information over-squashing in language tasks. Advances in Neural Information Processing Systems, 37: 98111--98142

work page 2024

[5] [5]

Chen, N.; Liu, Z.; Hooi, B.; He, B.; Fathony, R.; Hu, J.; and Chen, J. 2024. Consistency training with learnable data augmentation for graph anomaly detection with limited supervision. In The twelfth international conference on learning representations

work page 2024

[6] [6]

Dou, Y.; Liu, Z.; Sun, L.; Deng, Y.; Peng, H.; and Yu, P. S. 2020. Enhancing graph neural network-based fraud detectors against camouflaged fraudsters. In Proceedings of the 29th ACM international conference on information & knowledge management, 315--324

work page 2020

[7] [7]

Duan, M.; He, D.; Zheng, T.; Jia, L.; Song, M.; Wang, X.; and Feng, Z. 2025. Global Attribute-Association Pattern Aggregation for Graph Fraud Detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, 11616--11624

work page 2025

[8] [8]

Duan, M.; Zheng, T.; Gao, Y.; Wang, G.; Feng, Z.; and Wang, X. 2024. Dga-gnn: Dynamic grouping aggregation gnn for fraud detection. In Proceedings of the AAAI conference on artificial intelligence, volume 38, 11820--11828

work page 2024