When One Model Is Not Enough

Twin Training for Prioritized Decisions

Abstract

Many decision systems contain a small subset of cases with disproportionately high impact. Optimizing a single global objective can dilute performance on these priority cases, yielding models that are average–good but priority–weak. We introduce Twin Training, a practical framework that applies two sequential objectives to a single model: (i) a global phase that learns a stable decision function over the full distribution, followed by (ii) a priority-focused refinement phase that increases sensitivity within a predefined, high-impact subset of the feature space. At inference, a lightweight controlled arbitration mechanism applies a bounded rank-level adjustment for priority instances, preserving overall calibration and global stability while surfacing gains where they matter.

We evaluate Twin Training on public tabular benchmarks across four application areas: credit risk (German Credit, HELOC), income prediction (Adult Census), customer churn (Telco), and healthcare (Diabetes/Heart Disease). Priority subsets are defined by feature-derived, domain-aligned rules without relabeling (e.g., high loan amount or long duration; rare education × occupation combinations; high charges with long tenure; elderly or multi-comorbidity). In the Adult Census setting, Twin Training elevates a priority set of 92 minority-slice individuals to the top of the decision list with ≈ 84.7% precision, illustrating how priority gains can be surfaced without degrading overall behavior. The framework is model-agnostic (neural and gradient-boosted trees), avoids global loss reweighting and ensembling, and integrates cleanly into production via short priority refinement and transparent, bounded arbitration.