FedGCD

← All research

FedGCD: Federated Learning Algorithm with GNN based Community Detection for Heterogeneous Data

Published Journal of Internet Computing and Services (JICS) · 2023
  • Community detection
  • Client graphs
  • Non-IID learning
GitHub Paper

Overview

FedGCD studies federated learning through the relationships among clients. Instead of assuming that one aggregation structure fits every local distribution, it represents clients as a graph and identifies communities with related characteristics. Community information then guides model aggregation under non-IID data.

Core method

  1. Represent client relationships. Construct a client graph that captures differences or relationships among local models.
  2. Identify communities. Use the paper’s graph-based community-detection framework to organize heterogeneous clients into related groups.
  3. Aggregate with membership information. Use community structure and membership scores to determine how local models collaborate.

The paper also introduces nonIIDness as a measure of statistical heterogeneity, connecting the evaluation to the degree of distribution mismatch across clients.

Framework

FedGCD framework showing client graph construction, community detection, and aggregation
FedGCD framework: client relationships inform community-based collaboration.
View the algorithm diagram Algorithm diagram for the FedGCD training and community-aggregation procedure

Research artifact

The GitHub repository contains FedGCD.py, an early research prototype with graph construction, NMF-based factorization, community assignment, and aggregation routines. The paper provides the full methodological and experimental context for this prototype.

한국어 요약

FedGCD는 연합학습 클라이언트의 관계를 그래프로 표현하고, 유사한 특성을 가진 클라이언트들을 커뮤니티로 구성하는 연구입니다. 커뮤니티 구조와 멤버십 정보를 모델 집계에 반영해 non-IID 데이터 환경에서의 협업 방식을 설계합니다. 논문에서는 데이터 이질성의 정도를 평가하기 위한 nonIIDness 지표도 다룹니다.

공개 저장소에는 그래프 구성, 행렬 분해, 커뮤니티 할당 및 집계 과정을 담은 초기 연구용 프로토타입이 있습니다. 논문의 설명과 공개 코드의 구현 범위는 구분해서 확인할 수 있도록 논문과 코드 링크를 함께 제공합니다.

Publication

Wooseok Shin and Jitae Shin. “FedGCD: Federated Learning Algorithm with GNN based Community Detection for Heterogeneous Data.” Journal of Internet Computing and Services, 24(6), 1–11, 2023. doi:10.7472/jksii.2023.24.6.1.

FedVar examines client-weight variation at aggregation time. FedHyDRA combines distribution divergence, graph embeddings, and soft memberships to model structured heterogeneity.

Updated: