Accelerating the search for antimicrobial peptides with Google Antigravity
deepmind.google·2h ago
Data Science
Towards Data Science·1d ago·5 min read
Voting on Features: Federated Feature Selection Without Sharing Data
TL;DR
Feature Election is a novel federated feature selection method that allows hospitals to collaboratively train models without sharing patient data. It enables each site to vote on feature importance, improving model efficiency while maintaining data privacy.
✦ Why It Matters
Implement Feature Election to enhance your federated learning models without compromising data privacy.
Key Takeaways
- Use Feature Election for federated learning without data sharing.
- Each site can apply its trusted feature selector.
- Aggregate feature votes to improve model efficiency.
- Maintain patient privacy while training collaborative models.
Full Summary
In horizontal federated learning, multiple sites train a shared model without exchanging data, complicating feature selection. Traditional methods like Lasso or random forests require pooled data, which is unavailable in this context.
Feature Election addresses this by allowing each site to use its preferred feature selector, generating a binary vote and a score for each feature. The server then aggregates these votes, applying a tunable parameter called freedom degree to determine which features are included in the final model.
This method was implemented in NVIDIA FLARE and Flower, demonstrating improved model size and F1 scores. The approach ensures that no data leaves the local sites during the selection process, preserving patient privacy while enhancing model performance.