RegionFed: Federated Learning for Personalized Query Understanding in Heterogeneous Retail Environments
Regional search queries in retail environments inherently diverge across geographies, yet standard personalized federated learning often collapses on modern transformers due to tied embeddings and LayerNorm dynamics. RegionFed bypasses parameter-level interventions by operating entirely on gradient-level divergence. By diagnosing $\ell_2$ conflict between regional and global updates, the framework adaptively routes each region to appropriate personalization strategies without altering model architectures. Across benchmarks on T5 and RoBERTa, RegionFed-Meta achieved 92.27% accuracy under $(\epsilon \approx 0.60)$ differential privacy, matching the centralized training baseline.
Why it's worth reading
It addresses a known failure mode of personalized federated learning on modern transformers, presenting an architecture-agnostic gradient routing mechanism that closes the accuracy gap to centralized retail search baselines.