Positive-Unlabeled Graph Learning for Financial Crime Detection : Using Graph Neural Networks to Identify Financial Crime in Sparsely-Labeled, Real-World Data

Sammanfattning

Financial crime is a growing concern in today's digital financial systems, and traditional detection methods often crucially miss complex and interconnected patterns of illicit activity. By modeling financial network data as a graph, one can take advantage of relationships in the network such as corporate roles and invoice transactions, with graph neural networks (GNNs) being a strong architectural choice to learn patterns from them. However, a significant challenge in applying these models to real-world financial crime data is label sparsity and class imbalance, where confirmed cases are rare and the majority of cases are uninvestigated but not confirmed to be innocent. This complicates the training of these models and requires specialized learning strategies. In this thesis, we design and implement an end-to-end graph-learning pipeline for large-scale, real-world financial network data, with the purpose of detecting potential criminal activity. Our dataset contains approximately 2M organizations, of which only 1,780 are confirmed cases of financial crime, corresponding to 0.089% of the total. To address the issue of severe label sparsity, we apply a positive-unlabeled (PU) learning approach, designed for settings where only positive cases are confirmed while the rest are unclassified. We test the GNN models GAT, GATv2, GraphSAGE, and HAN, alongside an MLP as a performance baseline, and perform complementary explainability analysis to provide insight into the models' behavior. First, we conclude that PU learning in the form of PU loss, combined with stratified mini-batch sampling and focal modulation, is a fitting training strategy when working with extreme label sparsity under heavy class imbalance. Second, our results show that GAT was the most suitable GNN model in our PU learning setting, outperforming the other selected models in 4 out of 5 evaluation metrics. Notably, on the test set, GAT managed to concentrate known positives toward the top of its confidence ranking 57 times better than what random guessing would achieve. Third, explainability analysis on our dataset shows that high-confidence predictions of criminality are associated with company director composition, company maturity, and incoming invoices, whereas low-confidence predictions also depend on industry classification features.