# RCC-PFL: Robust Client Clustering under Noisy Labels in Personalized Federated Learning

**Authors:** Abdulmoneam Ali, Ahmed Arafa  
**arXiv:** [2503.19886](https://arxiv.org/abs/2503.19886)  
**Date:** 2025-03-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2503.19886v1/

## Key points

- Authors Abdulmoneam Ali and Ahmed Arafa address the challenge of accurately estimating cluster identities in personalized federated learning (PFL)
- PFL involves training different personal models and relies on clustering users into groups with similar objectives
- Noisy labeled data can lead to misleading loss function values and ineffective clustering
- The authors propose RCC-PFL, a label-agnostic data similarity-based clustering algorithm with three key advantages:
- Independently estimates cluster identities from training labels
- Acts as a one-shot clustering method before training
- Requires fewer communication rounds and less computation compared to iterative-based methods
- Validation using diverse models and datasets shows superiority over multiple baselines in terms of average accuracy and variance reduction
- RCC-PFL offers a robust solution for accurate cluster identity estimation in PFL settings, enhancing efficiency and effectiveness of personalized federated learning processes

## Summary

In their paper "RCC-PFL: Robust Client Clustering under Noisy Labels in Personalized Federated Learning," authors Abdulmoneam Ali and Ahmed Arafa tackle the challenge of accurately estimating cluster identities in a personalized federated learning (PFL) setting. PFL involves users training different personal models, with effective learning relying on clustering users into groups with similar objectives. However, noisy labeled data can lead to misleading values of loss functions and ineffective clustering. To overcome this challenge, the authors propose a label-agnostic data similarity-based clustering algorithm called RCC-PFL. This algorithm offers three key advantages: it independently estimates cluster identities from training labels, performs as a one-shot clustering method before training, and requires fewer communication rounds and less computation compared to iterative-based methods. The authors validate their approach using diverse models and datasets, demonstrating its superiority over multiple baselines in terms of average accuracy and variance reduction. This research provides valuable insights into addressing noisy labeled data in PFL settings through an innovative clustering approach. Their RCC-PFL algorithm offers a robust solution for accurate cluster identity estimation, enhancing the efficiency and effectiveness of personalized federated learning processes. Overall, this work contributes significantly to advancing the field of machine learning by improving cluster identification under challenging conditions and ultimately leading to enhanced model performance and user experience in personalized federated learning environments.