Layer-Wise Personalized Federated Learning for Mobile Traffic Prediction

With the evolution of mobile networks delivering high-performance network services to a myriad of devices, accurate mobile traffic prediction has become increasingly important. In recent years, federated learning (FL) has emerged as a communication-efficient approach, enabling collaborative model training without the centralized data aggregation. Despite its promising potential, FL-based mobile traffic prediction has following two major challenges: 1) data heterogeneity across regions: The diverse communication and mobility patterns inherent to different regions can lead to uneven traffic distribution. Training on such heterogeneous data can result in the global model failing to capture the unique patterns of specific regions, compromising consistent prediction performance across all regions; 2) communication efficiency concerns: The frequent exchange of large model weights during training leads to substantial signaling overhead in the FL. This added communication can pose a significant burden on the limited network bandwidth, potentially causing performance degradation in mobile networks. In this paper, we propose a novel personalized FL framework to address these challenges. Our framework enables a fine-grained federation through a layer-wise aggregation for the global model. This approach personalizes the global model to capture unique regional characteristics such as traffic spikes and other irregular patterns. In addition, we introduce an adaptive layer freezing mechanism to reduce communication costs during training. By selectively transmitting only the layers that require further training, our framework effectively enhances communication efficiency without sacrificing prediction performance. Extensive experiments on a real-world mobile traffic dataset demonstrate that our approach not only provides superior prediction accuracy compared to baselines but also achieves significant communication cost saving.