TL;DR: A new dataset, SynthFin-AML v10.0, addresses temporal leakage in GNNs by enforcing strict causal boundaries, preventing models from 'seeing' future edges during training.
Summary: Researchers released SynthFin-AML v10.0, a synthetic financial transaction dataset (100k nodes, 1.2M edges) designed to combat temporal leakage in Graph Neural Networks. It implements a 3-snapshot architecture with physically disjoint temporal windows for training, validation, and testing, ensuring GNNs only access information available up to a specific point in time. This prevents models from incorporating future edges into their embeddings, a common issue in dynamic graph evaluations.
Why it matters: AI builders working with GNNs on dynamic or time-series data should adopt rigorous temporal splitting strategies to avoid inflated performance. Evaluate your GNNs using datasets like SynthFin-AML to ensure your models learn from causally valid information and generalize effectively.
Source: reddit