Comparative Analysis of Federated and Centralized Learning Systems in Predicting Cellular Downlink Throughput Using CNN

Data traffic in cellular networks has surged due to the growing number of users and high-bandwidth applications. The quality of service (QoS) for users will degrade if the network resources cannot handle the increasing traffic volume. A user's application requires a minimum throughput to maintain QoS levels above the Service Level Agreement (SLA) promised by the operator. The network's ability to handle traffic growth will depend on well-prepared network resource management. KPI traffic prediction is one solution to anticipate traffic surges. This study utilized Federated Learning (FL) as a machine learning paradigm and employed CNN as a model to predict throughput in the cellular network. The efficacy of the model is compared with Centralized Learning (CL) and other deep learning models, including MLP, RNN, LSTM, and GRU. The experimental results indicate that the CNN model implemented in FL outperforms both CL and the other models. The number of rounds used in the FL system is held in two rounds, and the model's performance remains steady with an increasing number of clients, showing superior performance compared to CL.