Interpolation-Based Oversampling beyond Class Reweighting: An Exact Endpoint-Affine Selection-Curvature Decomposition
Abstract Interpolation-based oversampling changes both the total minority-class loss mass and the locations at which the loss is evaluated, so comparisons with an unweighted fit confound class reweighting with interpolation geometry. We construct a class-weighted control with exactly matched minority loss mass and objective normalization. Relative to this control, the residual gradient perturbation admits an exact endpoint-affine decomposition into an endpoint-selection term and an interpolation-curvature term. The decomposition yields explicit signatures for random oversampling, global interpolation, SMOTE, Borderline-SMOTE, and risk-weighted interpolation; for SMOTE, the selection term is governed by directed nearest-neighbour indegree imbalance. For smooth regularized empirical-risk minimization, the resulting signature determines the local optimizer response through the matched-control Hessian and yields a first-order, target-conditional validation score. Across 1,080 controlled comparisons on 12 imbalanced datasets using regularized logistic and probit regression, the full response attains a pooled median parameter-direction cosine of 0.997 and a median relative error of 0.119; learner-specific median relative errors are 0.115 and 0.122, respectively, while single-component ablations are materially less accurate under both learners. Across 216 learner-specific five-method held-out validation problems, the score recovers the oracle-best method in 87.5% of log-loss problems and 84.7% of Brier-score problems, with zero median selection regret. The framework isolates what interpolation contributes beyond matched class weighting and provides a mechanism-based approximation to candidate validation rankings without fitting every complete candidate model.
Authors
- Shiyou Sui (ORCID: https://orcid.org/0000-0001-9848-7352)
- Shengyu Chen
Institutions
- Tianjin University of Commerce (CN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-17
- DOI
- https://doi.org/10.21203/rs.3.rs-10773202/v1
- Primary Topic
- Imbalanced Data Classification Techniques
- Type
- preprint