| Data imbalance has been a persistent issue in machine learning. This is especially problematic in fraud detection, where fewer minority instances can have a significant effect on model performance. Several oversampling and undersampling techniques balance the data and enhance the model. In graph-based data, important information is encoded in the structure of edges and neighborhoods. This information cannot be obtained by conventional sampling techniques, which do not capture the graph relationships. The rapid growth of financial transactions creates the difficulty of handling huge, unbalanced data sets. In order to overcome this, we propose an approach that includes node and structural information by learning node embeddings that encode the graph’s neighborhood and structure. The node embeddings are learned using Graph Attention Networks (GAT), which are able to dynamically weigh neighbor nodes in a graph. We then employ a Variational Autoencoder (VAE) to oversample the minority instances based on the GAT embeddings to create synthetic data. The synthetic and the real samples are then used to train a Multi-Layer Perceptron (MLP) classifier in order to tackle class imbalance. Our experimental results indicate that this hybrid approach outperforms conventional oversampling techniques and offers a robust solution to a class imbalance for graph-structured data. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.