Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Rethinking Personalized Federated Learning with Clustering-based Dynamic Graph Propagation

Jiaqi Wang, Yuzhong Chen, Yuhang Wu, Mahashweta Das, Hao Yang, Fenglong Ma

TL;DR

This work tackles data heterogeneity in personalized federated learning by proposing FedCedar, a simple framework that clusters clients on the server, builds a dynamic weighted graph among cluster centers, and propagates knowledge across the graph to produce personalized models. The method integrates four components—client clustering, dynamic graph construction, graph-based knowledge propagation, and precise personalized model distribution—to capture hidden inter-cluster relations and improve local personalization without heavy tuning. Empirical results on MNIST, SVHN, and CIFAR-10 under IID and non-IID settings show FedCedar consistently outperforms a range of baselines, with notable gains on more complex non-IID datasets; ablation studies confirm the contribution of each module. A case study with synthetic topologies demonstrates the framework's ability to uncover and utilize hidden client relations, underscoring its practical impact for scalable, personalized FL.

Abstract

Most existing personalized federated learning approaches are based on intricate designs, which often require complex implementation and tuning. In order to address this limitation, we propose a simple yet effective personalized federated learning framework. Specifically, during each communication round, we group clients into multiple clusters based on their model training status and data distribution on the server side. We then consider each cluster center as a node equipped with model parameters and construct a graph that connects these nodes using weighted edges. Additionally, we update the model parameters at each node by propagating information across the entire graph. Subsequently, we design a precise personalized model distribution strategy to allow clients to obtain the most suitable model from the server side. We conduct experiments on three image benchmark datasets and create synthetic structured datasets with three types of typologies. Experimental results demonstrate the effectiveness of the proposed work.

Rethinking Personalized Federated Learning with Clustering-based Dynamic Graph Propagation

TL;DR

This work tackles data heterogeneity in personalized federated learning by proposing FedCedar, a simple framework that clusters clients on the server, builds a dynamic weighted graph among cluster centers, and propagates knowledge across the graph to produce personalized models. The method integrates four components—client clustering, dynamic graph construction, graph-based knowledge propagation, and precise personalized model distribution—to capture hidden inter-cluster relations and improve local personalization without heavy tuning. Empirical results on MNIST, SVHN, and CIFAR-10 under IID and non-IID settings show FedCedar consistently outperforms a range of baselines, with notable gains on more complex non-IID datasets; ablation studies confirm the contribution of each module. A case study with synthetic topologies demonstrates the framework's ability to uncover and utilize hidden client relations, underscoring its practical impact for scalable, personalized FL.

Abstract

Most existing personalized federated learning approaches are based on intricate designs, which often require complex implementation and tuning. In order to address this limitation, we propose a simple yet effective personalized federated learning framework. Specifically, during each communication round, we group clients into multiple clusters based on their model training status and data distribution on the server side. We then consider each cluster center as a node equipped with model parameters and construct a graph that connects these nodes using weighted edges. Additionally, we update the model parameters at each node by propagating information across the entire graph. Subsequently, we design a precise personalized model distribution strategy to allow clients to obtain the most suitable model from the server side. We conduct experiments on three image benchmark datasets and create synthetic structured datasets with three types of typologies. Experimental results demonstrate the effectiveness of the proposed work.
Paper Structure (22 sections, 2 equations, 4 figures, 2 tables)

This paper contains 22 sections, 2 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Model Overview
  5. Client Model Clustering
  6. Dynamic Weighted Graph Construction
  7. Knowledge Propagation and Aggregation
  8. Precise Personalized Model Distribution
  9. Experiment
  10. Experiment Setup
  11. Dataset Preparation
  12. Baselines
  13. Implementation Details
  14. Performance Evaluation
  15. Ablation Study
  16. ...and 7 more sections

Figures (4)

  • Figure 1: Overview of the proposed FedCedar ($K = 4$ as an example). Note that clients selected at time $t$ may be different from those that are selected at time $t+1$. Condition 1 means that the clients are selected at time $t$ as well as $(t+1)$, and Condition 2 means that the clients are not selected at time $t$.
  • Figure 2: Structured datasets with synthetic topologies.
  • Figure 3: Random index.
  • Figure 4: Hyperparameter study.