L2IR: Revealing Latent Intent in Graph Fraud Detection

Problem
This paper addresses the limitations of existing graph fraud detection methods, particularly those relying on Graph Neural Networks (GNNs), which struggle to identify fraudulent connections due to the dilution of fraud signals in heavily camouflaged environments. The authors highlight the inadequacy of current approaches in revealing the latent intent behind suspicious connections, a critical factor for effective detection. Additionally, the scarcity of annotated fraud samples poses a challenge for training robust detectors. This work is presented as a preprint and has not yet undergone peer review.

Method
The authors propose L2IR, a Latent Intent Revealing framework that leverages Large Language Models (LLMs) to extract intent-aware representations from user behaviors and suspicious connections. The architecture integrates a mechanism to uncover latent intent, allowing the model to differentiate between supportive and misleading links. The framework employs adaptive self-training to enhance its performance under limited supervision, effectively mitigating the challenges posed by the lack of annotated data. While specific details regarding the architecture and training compute are not disclosed, the method is designed to be compatible as a plug-in enhancement for existing GNN-based detectors.

Results
L2IR was evaluated on two real-world datasets characterized by pervasive camouflage, demonstrating significant improvements over strong baselines. The framework achieved an increase in Area Under the Precision-Recall Curve (AUPRC) by up to 8.27% compared to established GNN models. The results indicate that L2IR not only enhances detection accuracy but also provides a more nuanced understanding of the intent behind connections, which is crucial for distinguishing between benign and fraudulent interactions.

Limitations
The authors acknowledge that the reliance on LLMs may introduce challenges related to model interpretability and the potential for overfitting, especially in scenarios with limited annotated data. Additionally, the performance gains are contingent on the quality of the underlying GNN models, which may vary across different applications. The paper does not address the computational overhead introduced by integrating L2IR with existing GNN architectures, which could impact deployment in resource-constrained environments.

Why it matters
The implications of this work are significant for the field of fraud detection, particularly in graph-based contexts. By revealing latent intent, L2IR enhances the robustness and reliability of fraud detection systems, making them more effective in real-world applications where fraudsters employ sophisticated camouflage techniques. This framework not only contributes to the advancement of GNN-based detection methods but also opens avenues for further research into intent-driven approaches in various domains, including social network analysis and cybersecurity.

Authors: Jinsheng Guo, Zhenhao Weng, Yibo Liu, Yan Qiao, Meng Li
Source: arXiv:2605.26040
URL: https://arxiv.org/abs/2605.26040v1