Enhancing Fraud Detection with Graph-Based Machine Learning

Sammanfattning

Cryptocurrency fraud is a growing problem, driven by the growth of digital assets and the anonymity they provide, that affects both individuals and financial systems. Detecting fraud in this space is difficult because it often takes place within networks of transactions that traditional machine learning models struggle to represent. These models typically treat each transaction as an isolated event, ignoring the relationships between wallets and transactions and in turn, neglecting relationships that could be key to spotting illicit behaviour. Graph-based models, such as Graph Convolutional Networks (GCNs) offer a way to include this structure, yet there is limited research on how newer methods such as Heterogeneous Graph Convolutional Networks (HGCNs) compare to more established models. To address this, the thesis applies HGCNs and GCNs to the task of cryptocurrency fraud detection. Wallets and transactions are modelled as nodes in a heterogeneous graph, allowing the model to learn from both features and the network structure. The graph-based models are then compared against baseline models, Random Forest and Logistic Regression, on multiple fronts: implementation complexity, predictive performance, computational cost, and ability to handleimbalanced data. Results show that HGCNs provide stronger fraud detection performance in network-rich environments, especially when compared to GCNs. Random Forest remains the best overall model, however, these findings suggest that the inclusion of heterogeneous graphs can improve graph-based methods significantly. Opening the door to building more advanced surveillance tools within financial crime prevention