基于子群特征信息提取的个性化联邦学习方法
个性化联邦学习(Personalized Federated Learning, PFL)在联邦学习的基础上通过个性化建模的方式来解决数据异质性问题,在数据非独立同分布(Non-IID)场景下表现出较大的优势,推动了联邦学习在现实中的应用。然而,现有方法在客户端数据稀疏场景中正面临跨客户端知识协同中的特征匹配偏差的挑战。本文以提高客户端数据稀疏场景下个性化联邦学习的建模效果为目标,针对数据稀疏场景中局部特征高方差导致的客户端特征匹配可信度下降问题,提出了一种基于子群特征信息提取的PFL方法。该方法利用余弦相似性在稀疏高维空间中的鲁棒匹配特性构建相似度矩阵,辅助子群划分决策,最终提高了个性化模型的准确率。通过在数据量、异质性和数据集等多个维度上的实验验证,结果表明,所提出的方法相比于现有最优方法(SOTA),准确率提升了13.53\%,显著提高了客户端数据稀疏场景下个性化建模的效果。
Personalized Federated Learning (PFL) extends traditional federated learning by incorporating personalized modeling to address the challenges posed by data heterogeneity. It has demonstrated significant advantages in non-independent and identically distributed (Non-IID) scenarios, thereby facilitating the real-world adoption of federated learning. However, existing methods encounter challenges in client data-sparse settings, particularly in mitigating feature matching bias during cross-client knowledge collaboration. To enhance the modeling effectiveness of PFL in data-sparse scenarios, this study proposes a PFL method based on subgroup feature extraction to address the issue of reduced feature matching confidence caused by high variance in local features. The proposed method leverages cosine similarity’s robust matching properties in sparse high-dimensional spaces to construct a similarity matrix, which assists in subgroup partitioning decisions, ultimately improving the accuracy of personalized models. Experimental results across multiple dimensions, including data volume, data heterogeneity, and different datasets, demonstrate that the proposed method improves accuracy by 13.53\% compared to the state-of-the-art (SOTA) approaches, significantly enhancing the effectiveness of personalized modeling in data-sparse client scenarios.
蔡宣、周文安
北京邮电大学计算机学院,北京 100876,2022110782@bupt.cn 北京邮电大学计算机学院,北京 100876,zhouwa@bupt.edu.cn
计算技术、计算机技术
机器学习联邦学习个性化建模非独立同分布数据数据稀疏
Machine LearningFederated LearningPersonalized ModelingNon-IID DataData Sparsity
蔡宣,周文安.基于子群特征信息提取的个性化联邦学习方法[EB/OL].(2025-03-27)[2025-08-18].http://www.paper.edu.cn/releasepaper/content/202503-292.点此复制
评论