Supervised learning on graph-derived structural and weighted metrics for tax fraud detection

Structured tax fraud schemes frequently rely on complex relational arrangements designed to obscure beneficial ownership and coordinated economic behavior. This study investigates whether graph-derived structural and weighted descriptors can support supervised tax fraud detection in ego-centered fiscal relationship networks. Graph-level features capturing compactness, local cohesion, path organization, and weighted relational influence were extracted from institutional investigation graphs and evaluated using multiple supervised classifiers. Additionally, graph convolutional networks (GCN) and GraphSAGE (graph sample and aggregate) were evaluated as graph-native benchmark models operating directly on the original relational structures. Exploratory analyses revealed a dominant structural axis separating compact and distance-driven configurations between fraudulent and non-fraudulent networks. Among the evaluated models, XGBoost achieved the highest cross-validated F1-score and strong ROC–AUC discrimination, while GraphSAGE achieved competitive performance among graph-native approaches. Results suggest that interpretable graph-derived descriptors capture substantial predictive relational information embedded in the investigated fiscal networks. These findings highlight the potential of graph-based descriptors as auditable tools for tax fraud detection in institutionally sensitive environments.