RegionFed: Federated Learning for Personalized Query Understanding in Heterogeneous Retail Environments
RegionFed is a federated learning framework for retail search that operates at the gradient level, using the l2 conflict between regional and global gradients to diagnose heterogeneity, route regions to personalization strategies, and control personalization strength. It deploys on T5-Small, T5-3B, RoBERTa, and CNN with zero code changes, providing large gains on transformers where parameter-level personalization collapses below 10% accuracy on T5.
Retail search systems face data heterogeneity across regions. Standard federated learning sacrifices regional performance, while parameter-level personalization fails on transformers. RegionFed addresses this by operating at the gradient level, using gradient conflict to diagnose heterogeneity and adapt personalization. It is architecture-robust and works on T5-Small, T5-3B, RoBERTa, and CNN without code changes, achieving large gains on transformers.
RegionFed avoids parameter-level personalization failures on transformers by using gradient-level conflict signals. This approach treats models as differentiable black boxes, enabling deployment across architectures without code changes. The l2 conflict between regional and global gradients serves as a unified signal for heterogeneity diagnosis, routing, and personalization strength control.
Retail search systems can benefit from federated learning that preserves regional performance without sacrificing privacy. RegionFed's architecture-robust approach may lower barriers to adopting personalized federated learning in production, especially for transformer-based models.
RegionFed could enable retail search providers to improve query understanding and product recommendations across regions while maintaining data privacy and reducing model maintenance overhead due to its zero-code-change deployment.
Next signals include empirical validation on real retail search data, comparisons with other personalized FL methods, and potential adoption by e-commerce platforms seeking privacy-preserving regional personalization.