CS-pFedTM: Communication-Efficient and Similarity-based Personalised Federated Learning with Tsetlin Machine

TMLR Paper7355 Authors

Abstract: Federated Learning has emerged as a promising framework for privacy-preserving collaborative model training across decentralised data sources. However, data heterogeneity remains a major challenge, adversely affecting both the performance and efficiency of FL systems. To address this issue, we propose CS-pFedTM (Communication-Efficient and Similarity-based Personalised Federated Learning with Tsetlin Machine), a method that jointly incorporates communication-aware resource allocation and heterogeneity-driven personalisation. CS-pFedTM enforces communication budget constraints through adaptive clause allocation and tailors personalisation by using similarity between clients’ model parameters as a proxy for data heterogeneity. To further enhance scalability, the proposed framework integrates confidence-based aggregation and class-specific weight masking. Extensive experiments show that CS-pFedTM achieves reductions in communication and runtime costs, with up to $1352\times$ and $210\times$ reductions in upload and download communication respectively, and at least $1.43\times$ improvements in runtime efficiency, while maintaining performance comparable to state-of-the-art personalised FL approaches.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Tian_Li1

Submission Number: 7355