Federated feature selection with false discovery rate control.

Area of Science:

• Statistics
• Machine Learning
• Data Science
• Bioinformatics
• Genomics

Background:

• Federated feature selection is crucial for analyzing distributed data without sharing individual-level information.
• Privacy concerns and data heterogeneity across sites pose significant challenges to traditional feature selection methods.
• Existing methods struggle with controlling statistical error rates in a federated setting.

Purpose of the Study:

• To develop a privacy-preserving federated feature selection framework.
• To simultaneously identify relevant features and control the false discovery rate (FDR) across distributed datasets.
• To address challenges posed by data heterogeneity and communication costs in federated learning.

Main Methods:

• Proposed Fed-false discovery rate (Fed-FDR) framework for federated feature selection.
• Utilizes lower-dimensional coefficient estimates to preserve privacy and reduce communication overhead.
• Employs a generalized mirror statistic at a coordinating center to identify important features.

Main Results:

• Fed-FDR effectively controls the false discovery rate (FDR) in federated settings.
• Achieves strong statistical power in simulation studies, demonstrating high performance.
• Empirical studies confirm the method's validity and readiness for implementation.

Conclusions:

• Fed-FDR offers a robust, computationally efficient, and privacy-preserving solution for federated feature selection.
• The framework is resilient to heterogeneity in feature distributions and model parameters across sites.
• Demonstrates practical applicability and effectiveness for real-world distributed data analysis.