Distributed Core Network Traffic Prediction Architecture Based on Vertical Federated Learning

Network traffic prediction has always been an important research topic, frequently employed in intelligent network operations for load awareness, re-source management, and predictive control. Most existing methods adopt a centralized training and deployment approach, neglecting the involvement of multiple parties in the prediction process and the potential for training prediction models using distributed methods. This study introduces a novel wireless traffic prediction framework based on split learning, addressing the limitations of existing centralized methods. The proposed framework enables multiple edge clients to collaboratively train high-quality prediction models without transmitting large amounts of data, thus mitigating latency and privacy concerns. Each participant trains a dimension-specific prediction model using its local data, which are then aggregated through a collaborative interaction process. A partially global model is trained and shared among clients to tackle statistical heterogeneity challenges. Experimental results on real-world wireless traffic datasets demonstrate that our approach outperforms state-of-the-art methods, showing its potential and accuracy in Internet traffic prediction.