How effective are Graph Neural Networks in Fraud Detection for Network Data?

Graph-based Neural Networks (GNNs) are recent models created for learning representations of nodes (and graphs), which have achieved promising results when detecting patterns that occur in large-scale data relating different entities. Among these patterns, financial fraud stands out for its socioeconomic relevance and for presenting particular challenges, such as the extreme imbalance between the positive (fraud) and negative (legitimate transactions) classes, and the concept drift (i.e., statistical properties of the data change over time). Since GNNs are based on message propagation, the representation of a node is strongly impacted by its neighbors and by the network's hubs, amplifying the imbalance effects. Recent works attempt to adapt undersampling and oversampling strategies for GNNs in order to mitigate this effect without, however, accounting for concept drift. In this work, we conduct experiments to evaluate existing techniques for detecting network fraud, considering the two previous challenges. For this, we use real data sets, complemented by synthetic data created from a new methodology introduced here. Based on this analysis, we propose a series of improvement points that should be investigated in future research.