MSA-Fed: Model Similarity Aware Federated Learning for Data Heterogeneous QoS Prediction

Abstract

In the era of big data, QoS prediction is crucial for providing high-quality cloud services. However, conventional centralized approaches may pose privacy risks as they require users to upload QoS data. Additionally, variations in geographic and network environments can lead to QoS data heterogeneity, making it difficult to achieve learning efficiency with traditional methods. To address the privacy and heterogeneity issues, we propose a novel federated matrix factorization method with model similarity awareness for QoS prediction, called MSA-Fed. MSA-Fed clusters the local models uploaded by users during the learning process and performs differential aggregation and assignments of global models based on the clustering results. We evaluated the proposed framework on a publicly available and widely used real-world QoS dataset, and the experimental results demonstrate the effectiveness of MSA-Fed in achieving accurate QoS prediction, improving communication efficiency and maintaining users’ privacy.