Adaptive metagraph neural network assisted by metagraph search for financial fraud detection

Abstract

Financial transaction fraud detection is a critical technology for ensuring the security and stability of financial markets. Artificial intelligence, particularly graph neural networks, has demonstrated superior performance in fraud detection. However, challenges remain, such as limited interpretability, difficulty in adapting to new types of fraud in a timely manner, and incomplete data mining. To address these challenges, we propose a novel graph neural network model called Metagraph Fraud Detection Graph Neural Networks (MetaFraud-GNN), which leverages metagraph search and neural architecture search (NAS) techniques to automatically optimize the network structure for financial transaction fraud detection. MetaFraud-GNN extracts complex patterns from financial transaction networks through metagraph search algorithms, a technique that automatically mines key subgraph patterns. These metagraph capture fraudulent patterns and enable the model to more comprehensively uncover hidden information within transaction data, thus enhancing its processing efficiency. Additionally, the metagraph decoding algorithm optimizes the graph neural network structure by training on the most effective metagraph to adapt to evolving fraudulent methods. This approach improves both the accuracy and adaptability of fraud detection. We conduct experiments on three real-world public benchmark datasets YelpChi, Amazon, and Elliptic and demonstrate that our model significantly outperforms existing benchmark methods on various performance metrics. Such as F1-macro, Area Under the Receiver Operating Characteristic Curve(AUC) and Geometric Mean(GMean), by 4.46%, 2.67%, and 8.59% on YelpChi, 0.14% and 2.19% on Amazon,The F1 indicator has not been upgraded, respectively.