Evolutionary-Optimized Graph Neural Network for Fraud Detection with Topological Obfuscation
提出EOGFD模型，通过beta小波滤波器组分解图信号，并用遗传算法动态加权，解决拓扑混淆下欺诈节点与良性节点密集连接而欺诈节点间稀疏长距连接的检测难题，在四个数据集上AUC和F1-Macro分别提升10.49%和14.54%。
Graph fraud detection (GFD) is a challenging semi-supervised binary classification task that leverages graph neural networks (GNNs) to identify fraud nodes. However, most existing spectral GNNs rely on a strong assumption: that neighbor influences decay monotonically with increasing topological distance. This assumption significantly limits their effectiveness in scenarios involving topological obfuscation, where fraud nodes are densely connected to benign ones while maintaining sparse, long-range links with other fraud nodes. To address this challenge, we propose EOGFD, an Evolutionary-Optimized Graph neural network tailored for Fraud Detection under topological obfuscation. EOGFD first introduces a beta-wavelet band-pass filter bank that decomposes graph signals into multiple frequency components. Each component captures information from neighbors at different topological distances, enabling a rich and diverse set of graph representations. Building on this, we design a genetic algorithm-based dynamic coefficient optimization module that adaptively assigns weights to these components. This allows EOGFD to dynamically balance the influence of local and long-range neighbors, a critical capability for detecting obfuscated fraud patterns that violate the monotonic influence assumption. To further enhance generalization, we incorporate an early-learning guided regularization strategy that monitors training dynamics to mitigate overfitting to specific obfuscation patterns. This encourages the model to learn more robust and generalizable features capable of identifying a wider spectrum of fraud and benign topological structures. Extensive experiments on four real-world GFD datasets demonstrate EOGFD’s significant superiority over state-of-the-art methods, achieving substantial average improvements of 10.49% in AUC and 14.54% in F1-Macro. The code is available at https://github.com/BERA-wx/EOGFD.