FedU: Uniform Aggregation for Federated CNN-Based Downlink Throughput Prediction in Urban Area Cellular Networks

Abstract

One of the main challenges in implementing federated learning (FL) is the non-independent and identically distributed (non-IID) nature of inter-client data, which can affect server aggregation. When using a basic aggregation method such as FedAvg, the aggregation result is influenced by the client with the largest number of samples, potentially leading the resulting global model to fail to reflect the data characteristics of most clients. This study proposes the FedU method, which applies uniform aggregation weights to all clients and uses the total number of active clients as a coefficient weight. The CNN model was used to predict downlink throughput using a cellular network key performance indicator (KPI) traffic dataset collected from six base stations (BSs) within the same cluster, located in the urban area of Bandung, Indonesia. The performance of FedU was compared with three baseline methods, namely FedAvg, FedProx, and FedLBW. The experimental results show that the CNN model’s performance is no longer influenced by the number of samples, but is instead determined by the sample values and the dataset’s distribution. FedU achieves better CNN model performance than the three baseline aggregation methods across the evaluation metrics of loss, MAE, and $R^{2}$ .