Multi-Task Learning with Summary Statistics.

Area of Science:

• Machine Learning
• Statistical Learning
• Computational Biology

Background:

• Multi-task learning (MTL) integrates data from multiple sources to enhance model performance.
• Real-world MTL applications, particularly in healthcare, face significant data-sharing constraints.
• Existing MTL methods often require direct access to raw data, limiting their applicability.

Purpose of the Study:

• To develop a flexible MTL framework that utilizes summary statistics instead of raw data.
• To introduce an adaptive parameter selection method for MTL when only summary statistics are available.
• To provide theoretical guarantees and demonstrate practical utility in fields like genetic risk prediction.

Main Methods:

• Proposed a novel multi-task learning framework designed to work with summary statistics from distributed data sources.
• Developed an adaptive parameter selection approach using a variant of Lepski's method for data-driven tuning.
• Conducted a systematic non-asymptotic theoretical analysis of the proposed methods' performance.

Main Results:

• The proposed framework effectively handles data-sharing constraints by leveraging summary statistics.
• The adaptive parameter selection method allows for robust tuning without raw data access.
• Theoretical analysis provides performance characterizations under varying sample complexity and task overlap conditions.
• Extensive simulations validate the theoretical findings and demonstrate the method's practical performance.

Conclusions:

• The developed MTL framework offers a flexible and practical solution for training related models across domains with data-sharing limitations.
• This approach has significant potential for applications in areas such as genetic risk prediction and other fields requiring distributed data analysis.