Research map / Federated learning and privacy / Personalized and heterogeneous FL

Adaptive Group Personalization for Federated Mutual Transfer Learning

ICML 2024 · Federated learning and privacy › Personalized and heterogeneous FL · approach: Non-IID federated learning

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes AdaGrP, a tuning-free federated mutual transfer learning method that uses a nonparametric adaptive threshold correction to recover learnability (subgroup) structure under concept drift, with group-personalized aggregation.

Problem

Federated mutual transfer learning must recover unknown domain subgroup (learnability) structure and estimate parameters under communication and privacy constraints and concept drift, but existing clustered FL methods lack recovery guarantees and need hyperparameter tuning, while centralized DiffS u

Approach

Replace DiffS's fixed chi-square threshold with an adaptive threshold estimated per communication round by nonparametric kNN density estimation on standardized domain distances, and combine it with group-personalized FedAvg-style aggregation (global parameters averaged, heterogeneous parameters averaged per subgroup).

Key results

• AdaGrP achieves about 96.32% learnability structure recovery accuracy versus 91.52% for the best baseline on synthetic data.
• The paper reports a 16.9% average improvement in learnability structure recovery over state-of-the-art CFL baselines.
• On NOAA nClimDiv, AdaGrP yields a learnability structure similar to IECC climate zones, while AdaGrP (w/o) gives unbalanced clusters and FedDrift merges subgroups.

Measured effects reported in the paper

Surprising findings

• FedDrift produced about 11.8 clusters on average versus the true 3 and showed lower error at p=q=50 despite poor structure recovery.

Limitations

• Theoretical guarantees rely on assumptions (subgroup differentiation, known sigma_u and sigma_epsilon) and tightened conditions under finite local steps.
• Baselines IFCA and FeSEM were given the oracle number of clusters, and FedDrift required manual tuning of delta.

Experimental setup

Datasets: synthetic mixed-effects data NOAA nClimDiv

Benchmarks: synthetic mixed-effects data with concept drift NOAA nClimDiv average temperature prediction

Metrics: NMI RMSE mean squared prediction error

Baselines: IFCA FeSEM FedDrift AdaGrP (w/o) fixed threshold

Builds on: DiffS Clustered Federated Learning IFCA FedAvg Group personalized FL

Closest papers by meaning

• Decentralized Directed Collaboration for Personalized Federated Learning (CVPR 2024 · similarity 0.85)
• FedGCR: Achieving Performance and Fairness for Federated Learning with Distinct Client Types via Group Customization and Reweighting (AAAI 2024 · similarity 0.84)
• Clustered Federated Learning via Gradient-based Partitioning (ICML 2024 · similarity 0.84)
• Understanding the Statistical Accuracy-Communication Trade-off in Personalized Federated Learning with Minimax Guarantees (ICML 2025 · similarity 0.84)
• FedGMKD: An Efficient Prototype Federated Learning Framework through Knowledge Distillation and Discrepancy-Aware Aggregation (NeurIPS 2024 · similarity 0.84)
• Adaptive Federated Learning with Auto-Tuned Clients (ICLR 2024 · similarity 0.84)