Equally accurate models flag different people for the same review budget
Year: 2026
Type: preprint
Abstract: Risk models routinely rank individuals to fill fixed human-review budgets in advising, human resources, and credit underwriting, yet choosing among equally accurate models is rarely treated as consequential. We audit predictive and explanation multiplicity across five tabular benchmarks—employee attrition, higher-education student dropout, census income, consumer credit, and a redundant-signal synthetic task—over 30 repeated stratified splits. Within each split, every model within 0.02 validation AUC of the optimum fills the identical top-k review capacity. Near-optimal models differ by only 0.013–0.035 in test AUC and 1.1–6.7 percentage points in event capture, yet 44%–75% of individuals flagged at a 10% budget are omitted by competing near-optimal models. Conflict concentrates tightly in Quintile 1 around the capacity cut-off and persists inside explanations: unanimously flagged cases at a 20% budget trigger conflicting top occlusion reasons in 13%–84% of cases and divergent routing in 0%–34%. Selective consensus gating lowers routing instability by 29%–79%, primarily by withholding borderline cut-off cases and shrinking queue coverage. Under review constraints, selecting among equally accurate models is an allocative governance choice rather than a neutral technicality.
Author: Liu, Kang
Institution: Shandong Normal University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7235441461
