---
title: LLM-Enhanced Heterogeneous Graph Learning with Business Semantics for Financial Fraud Detection
url: https://www.ml-quant.com/papers/arxiv/2610.35298/
site: ML-Quant (https://www.ml-quant.com)
updated: 2026-10-30
license: Summaries CC BY 4.0; links go to the original sources
index: https://www.ml-quant.com/llms.txt
identifier: arXiv:2610.35298
source_url: https://arxiv.org/abs/2610.35298
featured: 2026-10-30
citations: unknown
topic: Risk, Credit & Banking
---


# LLM-Enhanced Heterogeneous Graph Learning with Business Semantics for Financial Fraud Detection

Role-aware prompts and LLM-derived semantic enrichment of transaction graph attributes improve detection of fraud in heterogeneous card-merchant networks.

- Source: https://arxiv.org/abs/2610.35298
- Identifier: arXiv:2610.35298
- Released: 2026-10-27
- First featured: Quant Letter No. 137 (2026-10-30): https://www.ml-quant.com/issues/2026-10-30/
- Citations (Semantic Scholar): not tracked
- Published in: not yet
- Topic: Risk, Credit & Banking
- Authors: Yixuan Chen, Hongyu Zhan, Jie Sheng, Weiyu Han, Shuai Chen, Tianyi Zhang, Xiao Tan, Jun Xia

## Abstract (arXiv, CC0)

Financial fraud detection increasingly relies on heterogeneous transaction graphs, in which entities such as cards and merchants play distinct business roles and the risk meaning of an attribute depends on the entity's role: a high transaction volume may be routine for a merchant but unusual for a card. Heterogeneous graph neural networks preserve typed entities and relations but typically treat these attributes as role-agnostic inputs and leave their business meaning implicit. Meanwhile, most existing methods integrating large language models (LLMs) with graphs focus on homogeneous graphs or heterogeneous graphs with naturally available textual attributes, leaving their use on transaction graphs with structured attributes underexplored. To address this issue, we propose SemHGT, an LLM-enhanced framework that interprets structured attributes according to entity roles and makes the business semantics explicit in three stages: (1) role-aware prompts elicit risk-related business descriptions from node attributes, and a text encoder maps them to semantic vectors that enrich node representations; (2) semantic similarity guides multi-round sampling of target-specific path instances under transaction-to-transaction schemas; (3) schema-level tokens that combine local path evidence with global neighbor summaries are processed together with the target transaction by a unified transformer. Across three benchmark datasets, SemHGT leads the evaluated baselines on most dataset--metric comparisons. Further analyses provide additional evidence for the effectiveness of SemHGT's design and clarify how its semantic modeling and evidence selection contribute to fraud detection. Code is available at https://anonymous.4open.science/r/SemHGT-528A/.

## Related

- [Attention-based Dynamic Multilayer Graph Neural Networks for Loan Default Prediction](https://www.ml-quant.com/papers/arxiv/2402.00299/): A dynamic multilayer network model has been created for improved credit risk assessment, considering borrower connections and their evolution over time.
- [DefaultGNN: A Dual-Perspective GNN Framework for Predicting Corporate Default from Buyer-Seller Transaction Networks](https://www.ml-quant.com/papers/arxiv/2609.25542/): A dual-perspective graph neural network framework predicts corporate defaults from buyer-seller transaction networks, improving approval rates by 7-11 percentage points without increasing default risk.
- [Unleashing the power of text for credit default prediction: Comparing human-written and generative AI-refined texts](https://www.ml-quant.com/papers/arxiv/2503.18029/): The study shows that using AI language model, ChatGPT, in lending decisions can improve credit default predictions and increase profitability in finance.
- [Neural Networks for Insurance Pricing with Frequency and Severity Data: A Benchmark Study from Data Preprocessing to Technical Tariff](https://www.ml-quant.com/papers/doi/10-1080-10920277-2025-2451860/): The article discusses the application of deep learning in insurance pricing, comparing different models and offering a method to interpret neural network insights through generalized linear models.
- [An Explicit Scheme for Pathwise XVA Computations](https://www.ml-quant.com/papers/arxiv/2401.13314/): A new simulation/regression scheme for a type of anticipated BSDEs is introduced, using neural network least-squares and quantile regressions, showing better results in high-dimensional and hybrid market/default risks XVA use-case.
- [Neural Networks for Insurance Pricing with Frequency and Severity Data: A Benchmark Study from Data Preprocessing to Technical Tariff](https://www.ml-quant.com/papers/arxiv/2310.12671/): The research evaluates different models for insurance claim frequency and severity data, presenting combined actuarial neural networks (CANNs) that merge baseline predictions with a neural network correction, and recommends using global surrogate models for practical deployment.