Semi-supervised Triply Robust Inductive Transfer Learning.

Area of Science:

• Machine Learning
• Biostatistics
• Genomics

Background:

• Integrating heterogeneous data from label-rich and label-scarce populations is crucial for improving predictive model accuracy.
• Covariate shift and the use of unlabeled data present significant challenges in transfer learning.
• Existing methods like double robustness may not fully address these complexities, especially with model misspecification.

Purpose of the Study:

• To propose a Semi-supervised Triply Robust Inductive transFer LEarning (STRIFLE) approach.
• To effectively integrate heterogeneous source and target population data, including unlabeled data, for enhanced learning.
• To achieve 'triple robustness' against nuisance model misspecification and distribution shifts.

Main Methods:

• Employed two nuisance models: a density ratio model and an imputation model.
• Combined transfer learning with surrogate-assisted semi-supervised learning strategies.
• Addressed high-dimensional covariate shift settings.

Main Results:

• The STRIFLE estimator demonstrated triple robustness, performing well even with misspecified nuisance models.
• It partially utilizes source population data when similarities exist, outperforming target-only methods.
• Theoretical guarantees and simulation studies confirmed the estimator's desirable properties.

Conclusions:

• STRIFLE offers a robust framework for transfer learning with heterogeneous and unlabeled data.
• The method is particularly effective in scenarios with covariate shift.
• Applied STRIFLE to develop a Type II diabetes polygenic risk prediction model for African Americans, leveraging data from a European population.