FedPA: Property-aware federated learning for cellular traffic prediction

Cellular traffic prediction plays a critical role in congestion management and resource optimization in communication networks. Along this direction, Federated Learning (FL)-based cellular traffic prediction methods enable distributed training while keeping data local, eliminating the need for data exchange between base stations (BSs) and enhancing privacy protection. Nevertheless, the global prediction performance of existing FL-based cellular traffic prediction methods remains suboptimal. This is primarily because they tend to independently capture either the temporal or spatial relationships, failing to fully accommodate the heterogeneous spatio-temporal properties of the traffic data across BSs. Moreover, these methods struggle to capture the diverse traffic trend variation properties, highlighting the need for deeper exploration of these trend variations. To address these issues, we propose a novel property-aware FL-based cellular traffic prediction method, named FedPA, which efficiently accommodates the heterogeneous traffic properties of individual BSs without requiring any public historical dataset. Firstly, to accommodate the heterogeneous spatio-temporal properties, our method introduces a representation-enhanced client selection strategy that selects clients with representative spatio-temporal properties to mitigate the effects of spatio-temporal heterogeneity. Secondly, to capture traffic trend variation properties, we design a discrepancy-aware hierarchical aggregation strategy that identifies clients with different traffic trend variation properties and effectively fuses diverse traffic patterns.We conduct extensive experiments on three real-world datasets, where FedPA outperforms existing state-of-the-art FL baselines in most cases. Specifically, compared to the best-performing FL-based baseline, our method improves the average RMSE, MAE, and R²metrics by 8.0 %, 5.3 %, and 4.6 %, respectively.