Leveraging Autoformer Deep Learning for Real-Time Throughput Forecasting in Cellular Adaptive Video Streaming

As mobile internet usage continues to grow and applications demand higher bandwidth, ensuring a high Quality of Experience (QoE) for users has become increasingly crucial. For bandwidth hungry and low-latency applications like video streaming and video conferencing, the ability to predict future bandwidth enables these applications to optimize their adaption decisions, leading to substantial improvements in the achieved QoE. In this paper, we focus on accurate prediction of the network throughput to improve the QoE in the context of DASH applications. Specifically, we focus on prediction of realtime throughput across different mobility patterns using 4G dataset traces collected in Ireland. Therefore, we propose an innovative throughput prediction framework using the Autoformer deep learning model, which is specifically designed to capture long-term temporal dependencies more effectively. Autofomer models outperform the state-of-the-art LSTM-based models, achieving significant improvement in the prediction accuracy of more than 70 % in some mobility scenarios. Further, we propose replacing the traditional estimation module in DASH players with the Autoformer-based prediction model to enhance overall perceptual video quality.