SynthFin-AML benchmark exposes temporal leakage in GNN AML models

Glabmayt2075 · reddit · 2026-09-01

The authors identified widespread temporal leakage in standard GNN baselines for anti-money laundering, where models see future edges during training. They released SynthFin-AML v10.0, enforcing strict causal boundaries with a 3-snapshot architecture to prevent this. By also fixing tabular distribution leakage, benchmarks show GraphSAGE (0.881 PR-AUC) only slightly outperforms LightGBM (0.848), confirming that GNN gains are often artifacts. The benchmark is proposed for PyTorch Geometric.

Original post →

More from Research

Research channel →