acceptodds
Under review as a conference paper at ICLR 2027

Behavior-Grounded Semantic Enrichment for Financial Fraud Modeling and Reasoning

Abstract

In financial fraud detection, rich semantic context can provide important evidence for transaction behavior modeling and fraud reasoning. However, public real-world financial datasets often lack rich semantics due to privacy constraints. Consequently, synthetic datasets incorporate generated semantics, but at the cost of behavioral realism; textual descriptions for contextual reasoning remain scarce. We address this gap through a semantic enrichment framework grounded in original transaction behavior to simulate multimodal financial data. We (1) propose a multi-agent semantic enrichment framework that generates interpretable financial semantics grounded in transaction behavior through role-specialized agents and consistency refinement, and (2) newly contribute a valuable multimodal financial fraud dataset, MS-FFSD, enriched with structured semantics and textual semantics while preserving real-data-grounded transaction behavior. Furthermore, we systematically analyze the quality and utility of semantic enrichment. Results demonstrate statistical fidelity and framework generalizability, while showing that richer semantics benefit fraud modeling and context-aware LLM reasoning. Overall, this work advances multimodal financial fraud research and bridges emerging LLM and multi-agent capabilities with operational anti-fraud practice. The framework and dataset are released at https://anonymous.4open.science/r/MS-FFSD-D7FD.

Primary area:
datasets and benchmarks
Keywords:
Financial Fraud Detection, Multimodal Financial Dataset, Multi-Agent Systems, Large Language Models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_hBPNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper proposes a multi-agent semantic enrichment framework to address the lack of semantic information. The framework generates structured user and merchant features and descriptions from transaction behaviors and preserves the original transaction structure. Based on the framework, the authors introduce a new multimodal financial fraud dataset MS-FFSD. The experiments evaluate semantic quality through statistical distribution alignment, demonstrate the benefits of enriched semantics for fraud detection across multiple GNN models and provide qualitative examples of LLM-based contextual reasoning.

Critical strengths
  1. Fraud datasets are usually hard to obtain since most institutions can not release this kind of dataset due to compliance requirements. A new fraud dataset (generation method) somehow narrows this gap so that the academic community can engage more with the problem. Existing LLM work on fraud detection has proven the importance of learning from semantic information. This work is a timely complement to support the methodology research.
  2. Good engineering design. The authors consider multiple sides of a transaction, including the user, merchant, and transaction itself. The use of demographic information is a bonus since it is good for downstream use in response AI studies, but it is also important to consider how to protect PII and avoid data misuse.
  3. The paper is well-written and easy to follow and the codebase looks good to the reviewer.
Critical weaknesses
  1. As a paper submitted to the dataset & benchmark area, the quantitative evaluation is limited. The paper emphasizes that enriched semantics could be used for LLM-based financial fraud reasoning. However, the full manuscript provides only two qualitative examples without any quantitative evaluation. It remains unclear whether the generated semantics actually improve LLM reasoning capability/accuracy on fraud detection, especially given that there are already some works letting LLMs leverage semantic information for fraud detection. Quantitative comparisons are needed to support the claim.
  2. While introducing a new financial fraud dataset is valuable, the experiments provide limited insights into what we can actually learn from it. Beyond modest performance improvements across GNN models, it would be helpful to know when and why semantic enrichment helps fraud detection, and what practical implications these findings offer. I would expect a more insightful analysis that goes beyond benchmarking existing models and provides actionable findings to fight fraud.
  3. The paper aims to address semantic scarcity through behavior-grounded enrichment, but 69.56% of users have only a single transaction and are enhanced by deterministic templates. Does this make those examples degrade to the regular user simulator like previous work? If yes, this raises a concern about how much of the enriched semantics is actually grounded in observed behavior.
Questions for authors

Check the above weaknesses

  1. You use IEEE-CIS to simulate the temporal distributions. However, this dataset is based on US consumers, and the proposed dataset is based on Chinese consumers with different payment patterns. If such migration would lead to a shifted distribution and how to prevent that?
Reviewer_NurCOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2(+1)
Soundness
1
Read review
Paper summary

The paper proposes a framework for taking existing anonymized transaction datasets and enriching it with plausible synthetic natural language descriptions of transactions, users and merchants.

Critical strengths
  • It tackles a very relevant problem in the lack of quality public data for fraud detection
Critical weaknesses
  • The mathematical formulation is too vague and it is unclear how it connects to the concrete generation pipeline. The authors only specify generic combinations of objectives with formulas that are left for appendices or not specified at all (e.g., FTF_T which I can't find anywhere). Furthermore, there is no concrete algorithm to minimize the objective, only a heuristic sequence of steps to generate structured attributes. As such, the formalization appears to be a post-hoc attempt to lend theoretical rigor to an otherwise heuristic pipeline. In my opinion, the paper would be much better served with a more concrete description of its approach in the main text. Most of the actually useful information is relegated and spread out over multiple appendices.
  • The overall approach seems backwards to me. In my experience, as far as synthetic simulators go, we need better simulators that generate transactions according to consistent user and merchant profiles. An example of this type of approach would be [1]. Instead, this work tries to retrofit an existing synthetic dataset with user and merchant descriptions, without improving the quality of the transactions themselves.
  • This enrichment, for the purpose of creating a multi-modal, seems misguided. Financial institutions do not necessarily have this type of natural language descriptions of their users, much less of merchants that show up in payments. In my experience, data is much more likely to be collected in databases, in the structured format that is only used as an intermediate step in this framework and not directly provided as enriched features. I don't understand in what scenario would these textual descriptions naturally arise as an available free text feature, ready to be used in a fraud detection pipeline.
  • ROC-AUC and F1-macro are not good metrics to evaluate fraud detection performance. Fraud detection requires low FPRs, as such it is typically only interested in the (operationally relevant) very low FPR region of the ROC curve. A better metric, in the same spirit, would be a partial ROC AUC up to a fixed low FPR (e.g., 1%). Macro F1 can also hide poor performance in the minority fraud class. I am puzzled as to what lead the authors to choose this metric in a binary classification setting where positive class F1 would be the natural and obvious choice.

[1] Yuan et al. PersonaLedger: Generating Realistic Financial Transactions with Persona Conditioned LLMs and Rule Grounded Feedback. https://arxiv.org/abs/2601.03149

Questions for authors

Clarifications:

  1. What is the motivation behind enriching the transaction data with textual information rather than the structured attributes directly?
  2. Where would this textual description live and how would it be obtained for use in fraud detection?

Suggestions:

  1. What would be the fraud detection performance using the enriched attributes directly through simple categorical encodings?
  2. Replace ROC AUC and macro F1 metrics with more domain relevant metrics.
Reviewer_1VU5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes a framework for enriching financial transaction datasets with synthetic temporal, demographic, geographic, merchant, and textual semantics. The framework preserves the original transaction records and labels while introducing externally informed attributes and entity descriptions. The authors contribute MS-FFSD, constructed from S-FFSD, and evaluate enrichment through distributional comparisons, graph-based fraud detection, and qualitative LLM reasoning examples.

Critical strengths
  1. The paper addresses a clear gap in financial fraud research. Privacy constraints often leave public transaction datasets with limited interpretable context. Developing resources that combine transaction behavior with semantic information is relevant to both fraud detection and LLM-based analysis.
  2. The dataset contribution is potentially useful. MS-FFSD organizes transactions, user attributes, merchant attributes, and entity descriptions into a linked resource.
  3. The framework makes a reasonable effort to constrain generation. Preserving transaction structure, amounts, and labels provides a useful foundation. External statistical priors, category hierarchies, and cross-entity consistency constraints are more principled than unconstrained description generation. Evaluation across 16 graph detectors and two additional private datasets also provides breadth.
Critical weaknesses
  1. The validity of the generated semantics remains insufficiently established. The paper documents external knowledge sources and construction rules, but agreement with the distributions used to generate attributes primarily demonstrates that those rules were followed. It does not establish that a particular user’s assigned demographics or a merchant’s assigned business category are accurate or uniquely supported by transaction behavior. Multiple substantially different semantic assignments could satisfy similar distributional and consistency constraints. Independent validation, sensitivity analysis, and explicit treatment of uncertainty would strengthen the rationale for these assignments.
  2. Label-dependent enrichment creates a potential evaluation concern. Appendix B.2 explicitly uses different temporal construction rules for fraud-related transaction blocks, including fraud-specific temporal distributions and burst templates. The paper needs to clarify whether validation or test labels participate in this process. If held-out labels influence generated timestamps and subsequent semantic attributes, the resulting features could indirectly encode the prediction target. Excluding explicit risk terms from textual prompts would not by itself resolve this issue. A label-blind enrichment experiment, with a clearly specified boundary between dataset construction and downstream evaluation, is necessary to interpret the predictive gains.
  3. Applying the framework to two private datasets is useful, but improved detection performance does not establish that generated attributes faithfully represent the underlying users or merchants. If trustworthy contextual attributes are available internally, masking and reconstructing them would provide a more direct validation opportunity. Otherwise, these experiments mainly demonstrate that the generated representations are useful under the chosen evaluation protocol. An additional public dataset with known contextual fields would also improve reproducibility and clarify applicability beyond S-FFSD.
  4. The qualitative examples show that generated descriptions change an LLM’s interpretation, but they do not establish that the resulting explanations are correct. Because merchant semantics are themselves inferred partly from transaction behavior, an explanation could become a plausible narrative constructed to fit the same evidence. Expert assessment or evaluation against independently known merchant context would help distinguish useful contextualization from unsupported justification.
Questions for authors
  1. Could the private datasets support direct fidelity evaluation by withholding known user or merchant attributes and comparing them with the generated semantics? If such ground truth is unavailable, what evidence supports entity-level fidelity beyond matching aggregate priors?

  2. Beyond S-FFSD, can the framework enrich another public transaction dataset? What minimum information must remain interpretable, and which components become unreliable when feature meanings, entity identities, or temporal information are unavailable?

  3. How are fraud labels used during enrichment, particularly in temporal construction and fraud-aware preservation? Are held-out labels excluded, and do the detection improvements persist under label-blind generation?

  4. How would descriptions be generated and updated in practice for new users, changing merchant behavior, and short transaction histories? What are the generation and refresh costs, and how does the framework prevent uncertain synthetic context from producing misleading fraud explanations?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.