<![CDATA[Financial technology is a means of fraud detection, and financial institutions say that still accounts for one-third of more than $32 billion in worldwide annual fraud losses. Credit card fraud detection aims to identify fraudulent transactions in massive financial transaction data using suspicious patterns and GNNs are compelling methods of PAR for knowledge-driven machine learning, such as fraud detection. This study shows a novel graph-based proposed feature engineering framework called GRAN that can directly leverage topological features within the financial transaction networks to efficiently model relational patterns without direct computational cost involved in standard GNN. To this end, the proposed methodology mitigates critical issues with classical fraud detection systems through multi-scale temporal features, improved graph construction techniques, extensive network-aware feature engineering and an explainable ensemble approach. More specifically, the framework extracts four types of features: customer-level node attributes, merchant-specific properties, network-aware transactional patterns and temporal behavioral signatures. Performance was optimized using a weighted ensemble of Random Forest (50%), Gradient Boosting (30%), and Logistic Regression (20%). We experiment over two credit card transaction datasets and perform better than the state of the art with our proposed GNN-Inspired Ensemble model, achieving F1=0.9463, precision=0.9533 and AUC=0.9914. The highest F1-score of 0.9481 was recorded for the Decision Tree model and in overall supervised methods outperformed unsupervised techniques. Hence, the balanced sampling strategy is able to effectively mitigate class imbalance problems that are inherent in fraud detection datasets. We draw two conclusions from our results: first, graph-based feature engineering is an effective approach to model complex fraud patterns with minimal computational requirements, and secondly, it allows for a straightforward interpretation which is crucial to financial mental models.]]>
Copyrights © 2025
