國立成功大學 統計與資料科學學系

Transfer learning has become a powerful statistical framework for integrating information from multiple related sources to improve estimation and prediction. However, in practice, measurement error from imperfect instruments or noisy data collection can distort covariates and responses, leading to biased inference and loss of efficiency.

This talk will introduce new transfer learning strategies that address measurement error in both parametric and nonparametric models. We first develop a unified method for error-contaminated Poisson regression, where count responses may be measured with error. By leveraging auxiliary source data, we construct bias-corrected estimators and adopt a model averaging approach to improve prediction stability under model uncertainty. Building on this idea, we further extend transfer learning to nonparametric generalized additive models with noisy covariates and develop a tree-based boosting framework that performs informative source detection, variable selection, and measurement error correction simultaneously. Simulation and real-data analyses, including breast cancer and global mortality studies, demonstrate substantial gains in estimation accuracy and predictive performance.