Abstract
Traffic data in cellular networks has increased significantly due to the growing number of users and high-bandwidth applications. If this increase in traffic is not balanced with sufficient network resource readiness, it can lead to a degradation in the quality of service that users receive. The applications used by users require a specific minimum throughput to achieve optimal quality of service. A solution is needed to maintain this quality of service by predicting the throughput in advance, allowing it to anticipate traffic data surges at an early stage. This research introduces the federated learning (FL) paradigm in machine learning, where a convolutional neural network (CNN) is used to predict downlink throughput in cellular networks. The performance of this model is measured and compared with that of the centralized learning (CL) paradigm as a benchmark. The addition of the dropout regularization technique to the CNN model is expected to enhance the prediction throughput performance. The results of this study prove that federated learning provides better performance than CL when implementing a CNN model to predict downlink throughput. Increasing the number of clients in FL can degrade performance, and the use of five rounds is the optimal value because the model still manages to produce better results without adding a computational burden.