| Summary: | Representing time-based transactions as graphs offers insights for fraud detection. However, limited and confidential datasets pose challenges. To address this, the project aims to assess the quality of generative model by using the generated data in training the fraud classifier model. A generative approach using graph autoencoders is employed to augment data for fraud detection. Experiments utilize the Elliptic dataset, a Bitcoin transaction dataset with 49 connected component graphs representing confirmations at twoweek intervals. A custom variational graph autoencoder (VGAE) is developed, encountering challenges with the defined loss function, resulting in excessive edge reconstruction. Among experimented three graph neural network (GNN) models—Graph Convolutional Networks (GCN), Graph Attention Networks (GAT) and GATv2, it is found that none effectively detect unseen illicit transactions. While GAT and GATv2 outperform GCN, models trained on reconstructed data perform worse than those on purely original data. It is observed the consistent metric fluctuations among models regardless of data used during training. This indicates the VGAE captures underlying patterns. This study highlights challenges and suggests future research directions, including adversarially regularized graph autoencoders, alternative architectures, efficient loss functions, and ensemble methods.
|
|---|