Robust angle-based transfer learning in high dimensions.

Area of Science:

• Statistical genetics
• Machine learning
• Bioinformatics

Background:

• Transfer learning is crucial for improving model performance with scarce target data.
• Challenges arise in high-dimensional regression with limited data and heterogeneous source populations.
• Existing methods often require individual-level source data, which may not be available.

Purpose of the Study:

• To develop a novel transfer learning method for high-dimensional regression with limited target data.
• To address the challenge of heterogeneous source populations when only model parameters are available.
• To propose a method that mitigates negative transfer and adapts to target signal strength.

Main Methods:

• Proposed a novel angle-based transfer learning (angleTL) method using parameter estimates from pretrained source models.
• Extended angleTL to incorporate multiple source models with varying relevance.
• Utilized high-dimensional asymptotic analysis to understand transfer benefits.

Main Results:

• AngleTL unifies several benchmark methods and adapts to target signal strength.
• The method effectively mitigates negative transfer in the presence of population heterogeneity.
• High-dimensional analysis confirmed the superiority of angleTL over existing approaches.

Conclusions:

• AngleTL provides an effective solution for transfer learning in high-dimensional regression with limited and heterogeneous data.
• The method is feasible for transferring genetic risk prediction models across biobanks.
• Leveraging parameter estimates enables knowledge transfer without requiring individual-level source data.