Wireless Traffic Prediction with L2 Paradigm Optimized Federated Learning

Abstract

Decentralized wireless traffic prediction is crucial for intelligent communication systems. But, in the conventional federated learning framework, ensuring the global model’s performance depends on frequent local-server communications, leading to substantial communication costs. This paper introduces an L2 norm-optimized federated learning framework, namely FedL2. FedL2 develops an L2 norm optimization method and devises an L2 norm-based local model uploading decision mechanism and an adaptive aggregation strategy for wireless traffic prediction tasks. In the local model uploading decision mechanism, the cumulative difference between the local and global models is compared with the communication threshold, selecting the most valuable local models for upload to reduce the local-server communication frequency. Furthermore, in the adaptive aggregation strategy, the performance of the prediction model is enhanced by quantifying the contribution of local models to the global model and capturing the spatial correlations between local base stations. The effectiveness of FedL2 is demonstrated on two real-world datasets, showing that FedL2 outperforms benchmark algorithms in terms of predictive performance and communication efficiency.