Go to ICLR 2025 Conference homepageConditionally Adaptive Graph Attention Networks for Credit Card Fraud Detection
Fraudulent transactions have been on the rise, leading to significant financial losses annually. In credit card fraud detection (CCFD), various predictive models aim to mitigate these losses by assessing transaction risk. While GNN-based methods have been employed to capture spatio-temporal transaction features, they often suffer from oversmoothing as graph layers increase, causing fraudulent and legitimate transactions to become indistinguishable. Existing semi-supervised methods that mask some labels have not fully resolved this issue. To address this, we propose the Multi-head Attention Conditional Variational Autoencoder (Ma-CVAE), which leverages weight distributions from imbalanced datasets and the Gumbel softmax distribution to construct more diverse reconstructed features, reducing feature homogenization. Then, We utilize Temporal Graph Attention Networks (TGAT) with a Multi-Attention mechanism to model risk propagation among transactions. Finally, classification probabilities are mapped to risk scores via a Multi-Layer Perceptron (MLP). Our approach achieves state-of-the-art performance, improving AUC scores by 1.45%, 3.05%, and 0.83% on three semi-supervised datasets: FFSD, YelpChi, and Amazon, respectively.