ID2Pattern: Pretraining Generalizable Identity Behavior Patterns for Transaction Risk Foundation Models (arXiv (Cornell University))

Summary

• High-cardinality identity fields such as card, device and account IDs are rich in behavioral signals but remain underexploited in payment foundation models.p.1
• ID2Pattern encodes identities with a Local Anonymous ID (LAID) tokenizer and pretrains a multi-scale pattern codebook on unlabeled transaction sequences.p.1
• Experiments use a 100M-scale industrial payment-risk dataset plus public benchmarks.p.2
• On the industrial test set, ID2Pattern improves AUC-ROC by 4.13%, T1 by 22.64% and AddCB by 19.67% over the semantic-attribute-only baseline.p.5
• Adding a hashed identity encoding lowers AUC-ROC by 5.87% against the same baseline, while LAID raises it by 1.11%.p.5
• Replacing LAID with hash-based encoding causes the largest average drop in the ablation, reaching 44.3% on CCT.p.7

Key insights

1. The generalizable signal of identity fields lies not in specific values but in the identity behavior patterns they exhibit across a transaction sequence.p.1
2. Existing foundation models encode card and device IDs like any other categorical attribute, giving no special treatment to their growing vocabularies.p.2
3. Because all LAID channels are bounded by the sequence length, the LAID vocabulary stays fixed as new entities emerge.p.3
4. Unseen identity values receive the same LAID tokens as seen ones that show the same behavior.p.3
5. Every ablated component lowers PR-AUC on all four tested datasets, and LAID contributes the most.p.7
6. On the CCT case study, a city rare in the long term but surging recently, combined with an unseen merchant, signals fraud at 222× the base rate.p.7
7. Adding ID2Pattern to a GRU backbone improves PR-AUC on IBM AML by 49.8%.p.8