Capturing Higher-Order Fraud: Multi-view Continual Learning for Streaming Fraud Detection

High-order fraud is increasingly prevalent in transaction networks, where fraudulent behaviors only emerge after aggregating weak signals across distant users. Existing graph-based methods typically rely on stacking multiple aggregation layers to capture such high-order patterns. However, in Streaming Fraud Detection (SFD), our theoretical analysis reveals that multi-layer aggregation can amplify the embedding drift bound under continuous concept drift, potentially leading to unstable model updates and degraded fraud detection performance. To address this issue, we propose MC-SFD, a Multi-view hypergraph Continual learning model for Streaming Fraud Detection, which reduces embedding drift through shallow hypergraph aggregation and effectively captures evolving high-order fraud. MC-SFD incorporates an RL-based Adaptive Hypergraph Constructor (RL-AHC) to dynamically evolve hypergraph structures and a Multi-view Personalized Prompt Generator (MPPG) for efficient adaptation with minimal parameter updates. In addition, both RL-AHC and MPPG integrate LLMs for policy initialization and semantic enhancement, respectively. Experiments on four SFD datasets show that MC-SFD consistently outperforms 23 state-of-the-art baselines by 1.63% to 3.71% in AUC.