Large Language Model-Enhanced Mutual Node-Edge Reinforcement for Inductive Fraud Detection in Financial Systems

Financial fraud detection requires advanced pattern recognition capabilities to address evolving risks. While existing graph neural networks (GNNs) have improved relational modeling, they still face limitations in synthesizing multi-modal features and handling unseen fraud patterns. This paper proposes large language model-enhanced mutual node-edge reinforcement (LLM-EMNER) for inductive fraud detection in financial systems. Our framework first leverages an LLM to separately generate features for nodes and edges, drawing upon their respective descriptions. Then, we develop an edge-aware graph encoder that dynamically fuses node attributes with transaction edge features through (a) a cross-modal attention mechanism that establishes mutual reinforcement between user nodes and transaction edges, (b) adaptive neighbor selection based on temporal-semantic similarity, and (c) edge-projected feature aggregation that preserves structural patterns. The generated node embeddings are then fed into a softmax classifier fine-tuned with contrastive learning objectives, leveraging the superior capability of contrastive learning in identifying complex fraud patterns from heterogeneous features. The LLM-EMNER component specifically employs a dual-stream architecture where edge features guide node representation updates, while node states refine edge semantics, creating dynamic feature reinforcement. For inductive learning, we design a meta-optimization strategy that enables rapid adaptation to new transaction patterns. The experimental results demonstrate that our method outperforms the baseline GNN methods. The LLM integration provides an average F1-macro improvement of 8.1% in detecting emerging financial fraud patterns compared to the state-of-the-art BWGNN and F2GNN approaches, validating the effectiveness of combining structural graph learning with LLM techniques.