No Peeking Ahead: Time-Aware Graph Fraud Detection

Data leakage in machine learning models occurs when training data contains information that won't be available during inference, leading to misleadingly high performance. This problem is particularly subtle in graph-based models, where future nodes or edges can silently alter the graph's topology and propagate leaked information through methods like GNNs. To prevent this, a time-aware approach is proposed, which involves building a temporal graph where edges are timestamped to reflect when a connection becomes known. This method ensures that models are trained only on information available at a specific point in time, as demonstrated with an insurance fraud detection example using a GraphSAGE model.