Dual-Stream Fraud Detection Framework Using Behaviour Feature Engineering and Graph Neural Networks for Financial Transaction Analysis
Data distributions are very skewed in financial fraud detection, and fraud behaviours are constantly changing in digital transaction systems, making financial fraud detection a tougher task. In this paper, they present a multimodal fraud detection framework that combines the representation learning of the financial network structure via graph-based methods with the feature engineering of the transactions' behaviour based on transactional data. The framework is assessed with two large-scale datasets: the PaySim financial transaction data is synthetic, and the Elliptic Bitcoin dataset is a transaction graph structure. Classical machine learning algorithms such as Logistic Regression, Random Forest and XGBoost are utilised for behavioural modelling, while GraphSAGE-based Graph Neural Networks are utilised for learning structural dependencies in transaction graphs. According to the experimental results, the performance of XGBoost is good, with an ROC-AUC of 0.9571 and an F1 score of 0.6812 with the dataset of PaySim. Furthermore, the GraphSAGE model achieves better performance than traditional models with a ROC-AUC of 0.9633, F1 score of 0.7707 and PR-AUC of 0.8188 in detecting relational fraud and is considered effective in such patterns. The acquired framework is well designed to provide effective evaluation while eliminating leakage with suitable validation strategies and class-weighted learning to deal with imbalanced data. Overall, the results demonstrate that blending behavioural and graph-based learning does significantly improve fraud detection capabilities in a complex financial context.