Reproduce “Principled Federated Random Forests for Heterogeneous Data” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Rémi Khellaf, Erwan Scornet, Aurélien Bellet, Julie Josse
Researchers introduce FedForest, a federated Random Forest algorithm that accurately reproduces centralized Classification and Regression Tree (CART) split decisions by aggregating client-side quantile sketches and additive sufficient statistics. This method effectively handles various forms of data heterogeneity, achieving performance comparable to centralized Random Forests on diverse datasets.
Key insights
- Centralized CART split decisions can be closely approximated, and often exactly replicated, in a federated setting using aggregated quantile sketches for candidate generation and additive sufficient statistics for impurity evaluation.
- The use of additive sufficient statistics for impurity calculation is crucial for maintaining performance under data heterogeneity, significantly outperforming heuristic methods that average local impurity gains.
- Splitting on client indicators provides a robust mechanism for personalized predictions, allowing the model to adapt to varying feature-outcome relationships across clients (outcome shift) without direct data sharing.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.