A new clustered federated learning algorithm for heterogeneous data in high-precision wireless sensing

Area of Science:

• Computer Science
• Machine Learning
• Wireless Communication

Background:

• Federated learning (FL) faces challenges with heterogeneous data in wireless sensing.
• Existing FL algorithms struggle to effectively handle data variability across devices.

Purpose of the Study:

• To develop a clustering-based federated learning algorithm for wireless sensing environments.
• To address data heterogeneity using Kullback-Leibler (KL) divergence for improved model personalization and accuracy.

Main Methods:

• Applied Principal Component Analysis (PCA) for dimension reduction of high-dimensional heterogeneous data.
• Calculated KL divergence distances between clients for clustering, incorporating an average distance for aggregated clients.
• Conducted federated learning within clusters to generate personalized models using wireless datasets.

Main Results:

• Iterative reclustering and model updates were performed to optimize cluster numbers and recognition accuracy.
• The proposed KL divergence-based algorithm demonstrated superior recognition accuracy compared to existing methods.
• Personalized models were successfully obtained for clients within identified clusters.

Conclusions:

• The clustering-based federated learning approach effectively handles heterogeneous data in wireless sensing.
• KL divergence is a viable metric for client clustering in federated learning, leading to enhanced performance.
• The proposed algorithm offers a promising solution for improving personalized model accuracy in distributed learning environments.