Towards More Efficient Data Valuation in Healthcare Federated Learning using Ensembling

Area of Science:

• Machine Learning
• Artificial Intelligence
• Medical Imaging

Background:

• Federated Learning (FL) enables collaborative model training without data sharing across institutions.
• Unequal contributions (data volume, quality, diversity) from participants pose challenges in FL.
• Shapley Value (SV) is a standard for fairly attributing contributions, but exact computation is often infeasible.

Purpose of the Study:

• To propose an efficient method for computing Shapley Values (SV) in Federated Learning (FL) settings.
• To address the computational expense of exact SV calculation in FL, particularly in healthcare.
• To introduce SaFE (Shapley Value for Federated Learning using Ensembling) for accurate contribution valuation.

Main Methods:

• Developed SaFE, an efficient algorithm for Shapley Value (SV) computation tailored for Federated Learning (FL).
• Employed ensembling techniques within SaFE to approximate SVs.
• Evaluated SaFE's performance against existing SV approximation methods.

Main Results:

• SaFE computes Shapley Values (SVs) that closely approximate exact SVs.
• SaFE demonstrates superior performance compared to current SV approximation techniques in FL.
• The method is particularly effective in heterogeneous medical imaging FL settings.

Conclusions:

• SaFE offers an efficient and accurate solution for valuing participant contributions in Federated Learning (FL).
• This approach is crucial for multi-institutional collaborative learning, especially in medical imaging where data heterogeneity is common.
• Accurate data valuation via SaFE facilitates fair recognition of institutional contributions in FL.