As financial systems grow more complex and interconnected, traditional fraud detection methods struggle to keep pace with increasingly sophisticated attacks. Graph-based approaches have been explored with a focus on cross-user interactions. In this paper, we propose a graph-based approach that focuses on individual cardholder behaviour. Each cardholder is modelled as an isolated graph capturing personal spending patterns. By leveraging advances in Graph Neural Networks (GNNs), we adopt an intra-group graph formulation where edges are restricted within a single cardholder’s transaction history. We construct three complementary edge types: temporal edges linking sequential transactions, similarity edges connecting behaviourally similar transactions, and merchant-based edges capturing repeated interactions with the same merchant. To isolate the effect of graph construction, we use a controlled experimental setup with fixed model architecture, training procedure, and evaluation protocol. We test the proposed model on two public datasets: the Sparkov and the IBM credit card datasets. We find strong model performance when transaction histories are dense. On Sparkov dataset, the model achieves 0.909 (F1-score) and 0.992 (AUC), substantially outperforming prior published results on the same dataset. On IBM, where transaction histories are sparse, the model achieves 0.763 (F1-score) and 0.962 (AUC), which highlights the importance of graph connectivity.