Abstract
A system and method for identifying counterfactual transaction pairs in a payment artificial-intelligence system. An atomic embedding vector associated with a query transaction is selected, and an inference embedding vector index is queried to retrieve transaction records whose embeddings satisfy a similarity criterion. Outcome metadata associated with the retrieved records is compared with an outcome of the query transaction. A retrieved transaction having a divergent outcome is paired with the query transaction as a counterfactual transaction pair. An embedding distance between the pair members indicates decision-boundary proximity. The pair may be stored in a counterfactual-pair registry and used for decision boundary analysis, hard-negative training, robustness analysis, fairness auditing, or investigator support. A decision-boundary proximity signal may further be provided to an authorization rule engine to generate an ALLOW, CHALLENGE, or DECLINE authorization outcome. Pair-quality indicators may be evaluated to quarantine agent-synthesized transaction variation from a training corpus.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Kala, Durga S.; Pandey, Om; Barcelos, Tanner; Nativi, Benjamin; and Xu, Jun, "AUTOMATED COUNTERFACTUAL PAIR MINING IN PAYMENT ARTIFICIAL-INTELLIGENCE SYSTEMS", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12399