Á. Cía Mina, J. López Fidalgo, L. Deldossi, C. Tommasi
In large-scale regression problems, subsampling is often used to enhance computational efficiency. Traditional subsampling techniques primarily focus on accurate parameter estimation, yet in many practical applications, the ultimate objective is to improve predictive performance. This study presents a new subsampling strategy for linear models that explicitly accounts for model misspecification. The approach leverages the distribution of covariates and is particularly useful in scenarios where acquiring response variable labels is expensive. By targeting the reduction of bias in the random-X prediction error, the proposed method enhances predictive accuracy. Theoretical results establish its advantage in lowering prediction mean squared error, and simulation studies further validate its effectiveness compared to existing approaches.
Keywords: Model misspecification, Random-X regression, Optimal Design of Experiments
Scheduled
Design of Experiments I
June 11, 2025 3:30 PM
MR 1