Tax Fraud Detection Through Hybrid Relational and Anomaly Learning.

Tax fraud remains a significant challenge for revenue authorities as fraudulent activities increasingly involve coordinated networks of taxpayers, shell companies, and concealed financial transactions. Conventional detection methods typically rely on either relational learning or anomaly detection alone, making it difficult to capture both the structural relationships among taxpayers and unusual behavioral patterns within a single framework. This study presents a hybrid deep learning framework that integrates Graph Neural Networks (GNNs) with Autoencoders to improve tax fraud detection. The GNN component learns contextual representations from taxpayer interaction networks, capturing dependencies and hidden relationships among entities, while the Autoencoder models normal taxpayer behavior and flags deviations that may indicate fraud. An attention-based fusion mechanism combines the relational and behavioral representations before classification, allowing the model to exploit complementary information from both learning approaches. The framework was implemented and evaluated on a synthetic taxpayer dataset of 8,000 records (19% fraud prevalence) with a constructed relational graph of 31,905 edges, using stratified 70:15:15 partitioning and focal-loss training to address class imbalance. On the held-out test set, the framework achieved an accuracy of 0.8125, precision of 0.5049, recall of 0.6754, F1-score of 0.5779, and ROC-AUC of 0.8392, outperforming Logistic Regression, Random Forest, XGBoost, LightGBM, and a Multi-Layer Perceptron baseline, all of which recorded markedly lower recall 0.04–0.56 despite comparable accuracy. Bootstrap 95% confidence intervals confirmed the stability of this advantage and an ablation study showed each of the three information streams to be individually necessary, with the relational stream carrying the largest mean attention weight 0.457. Error and calibration analysis showed the model favouring detection sensitivity over false-alarm suppression, consistent with audit-prioritisation use. These findings indicate that combining relational and anomaly-based components yields measurable gains over conventional classifiers for imbalanced tax fraud detection.