An Adaptive Particle Swarm Optimization-Based Graph Neural Network (APSO-GNN) for Explainable Financial Fraud Detection
Abstract
Financial fraud detection has become increasingly challenging with the rapid growth of blockchain transactions, where transaction relationships naturally form complex graph structures. This paper presents APSO-GNN, a hybrid framework that integrates Adaptive Particle Swarm Optimization (APSO) with Graph Neural Networks (GNNs). The framework automatically optimizes model hyperparameters to improve fraud detection performance on highly imbalanced transaction graphs. Three representative GNN architectures, namely Graph Convolutional Networks (GCN), Graph Attention Networks (GAT), and GraphSAGE, are evaluated using the Elliptic Bitcoin dataset. The dataset comprises 203,769 transactions, of which only 2.8% correspond to fraudulent activities. Extensive experiments using five independent random seeds demonstrate that APSO consistently outperforms the corresponding baseline models. In particular, GraphSAGE-APSO achieves an AUC of 0.946 ± 0.006 and an F1-score of 0.493 ± 0.031, representing improvements of 18.4% and 181.7%, respectively, over the baseline (p < 0.001). Statistical analyses further confirm the robustness and reproducibility of the proposed optimization strategy across multiple experimental runs. To enhance model transparency, GNNExplainer is employed to identify influential graph structures and node features, revealing that fraud detection is primarily driven by local structural connectivity and a small subset of discriminative transaction attributes. These findings demonstrate that APSO-GNN provides an effective and interpretable solution for blockchain fraud detection, offering improved predictive performance while supporting trustworthy deployment in real-world financial systems.
