Paper Type
ERF
Abstract
Reject inference (RI) focuses on selective label observability in credit scoring by incorporating information about rejected applicants whose repayment outcomes are not observed under historical approval policies. We propose a theory-inspired design science artifact for RI by framing inclusion as an explicit exploration-exploitation problem: the system must explore beyond the accepted-only samples by selecting informative rejects, while exploiting by constraining distribution shift and stopping when marginal gains saturate. We operationalize this as a quadratic unconstrained binary optimization (QUBO) that rewards informativeness (model disagreement), penalizes redundancy (feature-space similarity), and encodes conservative composition controls (e.g., good:bad quotas and per-round caps) for iterative pseudo-label inclusion. We plan on evaluating the approach in an iterative RI pipeline using standard ensemble credit-risk learners on two settings, a public U.S. LendingClub dataset and a proprietary China micro-lending dataset, comparing against classical baselines. The QUBO representation separates the policy (optimization specification) from the solver (implementation choice). QUBO can be solved by classical, hybrid, or quantum-annealing solvers. This quantum-ready formulation aims to make the RI inclusion policy executable at scale.
Paper Number
1867
Recommended Citation
Ding, Wenwen and Sykes, Tracy Ann, "Exploration-Exploitation for Reject Inference: A Quantum-Ready Optimization Policy for Credit Decision Systems" (2026). AMCIS 2026 Proceedings. 16.
https://aisel.aisnet.org/amcis2026/ai_systdesign/ai_systdesign/16
Exploration-Exploitation for Reject Inference: A Quantum-Ready Optimization Policy for Credit Decision Systems
Reject inference (RI) focuses on selective label observability in credit scoring by incorporating information about rejected applicants whose repayment outcomes are not observed under historical approval policies. We propose a theory-inspired design science artifact for RI by framing inclusion as an explicit exploration-exploitation problem: the system must explore beyond the accepted-only samples by selecting informative rejects, while exploiting by constraining distribution shift and stopping when marginal gains saturate. We operationalize this as a quadratic unconstrained binary optimization (QUBO) that rewards informativeness (model disagreement), penalizes redundancy (feature-space similarity), and encodes conservative composition controls (e.g., good:bad quotas and per-round caps) for iterative pseudo-label inclusion. We plan on evaluating the approach in an iterative RI pipeline using standard ensemble credit-risk learners on two settings, a public U.S. LendingClub dataset and a proprietary China micro-lending dataset, comparing against classical baselines. The QUBO representation separates the policy (optimization specification) from the solver (implementation choice). QUBO can be solved by classical, hybrid, or quantum-annealing solvers. This quantum-ready formulation aims to make the RI inclusion policy executable at scale.
When commenting on articles, please be friendly, welcoming, respectful and abide by the AIS eLibrary Discussion Thread Code of Conduct posted here.
Comments
AI SYSTEM