Scholarlink

A grand challenge in personalised federated learning (PFL) is the data heterogeneity. Although various methods tried to address this challenge, they are either inefficient or show low generalisation capability. In this work, we propose a PFL framework One-shot Federated Clustering Hashing for Personalization (OFCHP) that addresses the efficiency and generalisation while considering the Non Identically Independent Distributed(Non-IID) scenario. In OFCHP, we adapt the efficient high-dimensional data search algorithm Locality Sensitive Hashing (LSH) to the PFL setting to achieve efficient client clustering. To ensure the stability of clustering for new clients, we resolve the two scenes under clients imbalance. One is that the clustering effect is not good due to insufficient client data in small datasets. Secondly, our technique could also alleviate the late client issue and allow late clients to find their clusters quickly. The experimental results show that our method outperforms traditional federated algorithms in terms of accuracy, generalisation and efficiency under a heterogeneous data environment.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Efficient Clustered Federated Learning by Locality Sensitive Hashing

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Lishan Yang,
  • Alireza Seyed Shakeri,
  • Liangxi Pu,
  • Weitong Chen,
  • Yanjun Shu

摘要

A grand challenge in personalised federated learning (PFL) is the data heterogeneity. Although various methods tried to address this challenge, they are either inefficient or show low generalisation capability. In this work, we propose a PFL framework One-shot Federated Clustering Hashing for Personalization (OFCHP) that addresses the efficiency and generalisation while considering the Non Identically Independent Distributed(Non-IID) scenario. In OFCHP, we adapt the efficient high-dimensional data search algorithm Locality Sensitive Hashing (LSH) to the PFL setting to achieve efficient client clustering. To ensure the stability of clustering for new clients, we resolve the two scenes under clients imbalance. One is that the clustering effect is not good due to insufficient client data in small datasets. Secondly, our technique could also alleviate the late client issue and allow late clients to find their clusters quickly. The experimental results show that our method outperforms traditional federated algorithms in terms of accuracy, generalisation and efficiency under a heterogeneous data environment.