Skip to content

Back to the catalog

RegionFed: Federated Learning for Personalized Query Understanding in Heterogeneous Retail Environments
arxiv.orgpaper

RegionFed: Federated Learning for Personalized Query Understanding in Heterogeneous Retail Environments

RegionFed is a federated learning framework for personalized retail search query understanding that operates at the gradient level rather than parameter level, using the ℓ2 conflict between regional and global gradients to diagnose heterogeneity and route regions to personalization strategies. Unlike existing parameter-level personalized FL methods that collapse below 10% accuracy on transformers like T5, RegionFed deploys unchanged across T5-Small, T5-3B, RoBERTa, and CNN architectures. Evaluated on Amazon ESCI, Amazon Reviews, and LEAF-FEMNIST, RegionFed-Meta reaches 92.27% accuracy, closing to within 0.23pp of a privacy-violating centralized upper bound while providing (ε≈0.60)-differential privacy.

TL;DR Gradient-level federated learning framework that avoids the transformer-collapse failure of parameter-level personalized FL, matching centralized accuracy with differential privacy.
Open original ↗
Document

What it claims

RegionFed adds detail the summary compresses: the reason parameter-level personalized FL collapses on transformers is specifically tied embeddings and LayerNorm interactions, which push T5 accuracy below 10%. By working only on gradients and treating each model as a differentiable black box, RegionFed sidesteps that failure and drops onto T5-Small, T5-3B, RoBERTa, and CNN with zero code changes.

How it works

The ℓ2 conflict between regional and global gradients is a single signal doing three jobs the summary only partly names: (i) diagnosing heterogeneity, (ii) routing each region to the cheapest sufficient personalization strategy, and (iii) adaptively controlling personalization strength. The third — a continuous strength knob, not just a strategy choice — is the piece the summary omits.

flowchart TD
    A[regional vs global gradient ℓ2 conflict] --> B{heterogeneity level}
    B -->|low| C[global model suffices]
    B -->|high| D[route to cheapest sufficient strategy]
    D --> E[adaptive strength control]
    C --> F[updated model]
    E --> F

Evidence

The centralized upper bound the 92.27% is measured against is named precisely: Centralized + Regional Weighting at 92.04%, making RegionFed-Meta's Δ=0.23pp gap fall within 1σ — i.e. statistically indistinguishable, not merely close. The abstract also states an O(1/√T) convergence guarantee alongside the (ε≈0.60)-DP claim, which the summary does not mention. Datasets and architectures are as summarized: Amazon ESCI, Amazon Reviews, LEAF-FEMNIST across four architectures.

Limits

This is an abstract-only arXiv page (submitted 4 Sep 2026); no code link, per-dataset or per-architecture breakdown, or ablation is visible. The headline 92.27% is a single aggregate — the abstract does not show which architectures or regions drive it, nor the CNN gains it calls merely "consistent." Anyone weighing adoption needs the paper for the routing thresholds, the strength-control mechanics, and how the ε≈0.60 budget interacts with the reported accuracy.

Organisations named here

  • Amazon ESCI
  • Amazon Reviews
  • LEAF-FEMNIST

Technologies named here

  • T5
  • RegionFed
  • T5-Small
  • T5-3B
  • RoBERTa
  • CNN
  • Transformer

Techniques named here

  • Gradient-level personalization
  • Parameter-level personalized federated learning
  • Differential privacy in federated learning

Related in research