Effect of Spatial, Temporal and Network Features on Uplink and Downlink Throughput Prediction

Recently, there have been many attempts to apply Machine Learning (ML)-based prediction mechanisms in wireless networks. One open question is how reliable such predictions can be, and how well ML models can learn from the radio environment. In this paper, we present initial results on Quality of Service (QoS) prediction using the example of throughput prediction. We focus on suggesting new sets of features that can improve the prediction performance for different prediction horizons. Thereby, we identify important features that have a large impact when using radio environment data as input for ML models. To this end, we consider information from space, time, and network domains. In particular, we show that features, such as cell throughput and previous users' data can significantly improve the ML model performance. Besides the importance of input features, we also investigate how the prediction performance deteriorates for different prediction horizons.