E-Commerce Transaction Fraud Detection Using GNN-Based Feature Embedding and Machine Learning Methods

E-commerce fraud poses a significant threat to digital economies, causing financial losses and weakening consumer trust in online platforms. Existing detection systems typically rely on supervised machine learning (ML) classifiers such as Artificial Neural Networks, Random Forest, and Gradient Boosting applied to tabular transaction features. ML classifiers process transactions independently and cannot capture relationships across shared entities such as devices, IP addresses, and payment methods. Other approaches use Graph Neural Networks to capture relational dependencies between transactions. End-to-end GNN classifiers present deployment challenges including high computational cost and limited interpretability. To address this limitation, this study proposes a hybrid fraud detection pipeline to extract relational embeddings from a transaction graph and combine them with tabular features for conventional machine learning classifiers. Experiments are conducted on the Kaggle Fraudulent E-commerce Transaction dataset, which contains 1,472,952 training records and 23,634 test records. Three classifiers, namely Artificial Neural Networks, Random Forest, and Gradient Boosting, are evaluated under three scenarios: tabular only, end-to-end GNN, and hybrid GNN embedding. Across all three classifiers, augmenting tabular features with relational embeddings yields consistent improvements over the tabular-only baseline. The Random Forest hybrid pipeline obtains the most balanced result (accuracy = 0.9688, F1=0.6091, PR-AUC = 0.6171), although the absolute gain over the best tabular baseline is modest, indicating that GNN-based feature embedding provides a complementary rather than transformative source of predictive signal.