Recently, the team of Professor Yuchen Zheng and Associate Professor Xiaofang Wang from the School of Information Science and Technology (School of Cyber Science and Technology) has made new progress in personalized federated learning. The paper titled CEHGLS: A Communication‑Efficient Head Gradient Linear Search for Personalized Federated Learning under Data Heterogeneity has been published in IEEE Internet of Things Journal (CAS Q1 Top, IF: 8.9).

In practical scenarios, federated learning inevitably suffers from data heterogeneity, which is mainly reflected in severe model aggregation bias. To improve the capability of conventional homogeneous federated learning methods against this issue, this paper proposes a modular aggregation‑bias correction strategy named CEHGLS, short for Communication‑Efficient Head Gradient Linear Search.

Drawing on the core idea of decoupling general representations and task‑specific heads in personalized federated learning, the algorithm takes mitigation of aggregation bias as an explicit optimization objective. Specifically, the designed Head Gradient Linear Search module (HGLS) fully explores the rich directional information contained in client head gradients. It performs active linear search in the parameter space to find the optimal gradient combination that maximizes local adaptability, so that local models can better adapt to their unique data‑distribution characteristics. To guarantee communication efficiency, the Head Gradient Compression module (HGC) is introduced to exchange correction signals via high‑compression‑ratio gradient transmission. CEHGLS can be flexibly plugged into existing classic federated‑learning frameworks and extends their capability for highly‑heterogeneous data environments.

Framework of the CEHGLS algorithm

This paper is co‑first‑authored by master students Zouquan Chen and Lidong Zheng. Professor Yuchen Zheng and Associate Professor Xiaofang Wang are the corresponding authors.