This thesis develops and evaluates two complementary approaches to test whether an unsupervised model can increase risk coverage. The models used is a supervised XGBoost classifier and an unsupervised Variational Graph Autoencoder (VGAE) with a GATv2 encoder. Both models are evaluated on a shared synthetic test set produced by the AMLGentex framework, consisting of 100,000 accounts observed over 50 time steps, containing 133,504 internal transactions and 279 SAR labelled accounts. The supervised model is trained on a separate 100 step training set with 80/20 train/validation split that prevents label leakage, and its hyperparameters are tuned to maximise the validation AUCPR. On the test set the model achieves an AUCPR of 0.587 and identifies 99 of 279 SAR accounts at precision 0.971 and recall 0.355. The unsupervised model is trained on temporal subgraphs of the same training data without using any labels. Anomaly scoring was based on per node KL divergence. The KL signal is reported at two minimum degree thresholds: at degree 10, every account in the top 10% of the resulting ranking carries a SAR label; at degree 3, the model covers 69.5% of the test SAR population and surfaces accounts whose structural position in the transaction graph deviates from the learned norm. The Spearman rank correlation between the two models’ scores is r = 0.165 across the shared scored population, confirming that the two models respond to different signals in the data.