LLM-GNN Soft Prompt Framework for Fraud Detection

The LGSPF (LLM-GNN Soft Prompt Framework) has been introduced as a solution for fraud detection, tackling the issue of insufficient textual data in graph-related tasks. This innovative approach employs soft prompts to connect graph structures with semantic spaces, removing the dependence on textual attributes. It features a parallel Graph Neural Network (GNN) encoder that converts multi-relational topologies into graph tokens, enhancing the understanding of fraud by large language models (LLMs). Additionally, the framework supports end-to-end optimization, addressing the challenges posed by hard prompts and effectively capturing intricate semantic details in multi-relational fraud detection.