M2ATURE: Mobile Multistage Throughput Prediction for Adaptive Video Streaming in Cellular Networks
Accurate Throughput Prediction (TP) represents a real challenge for reliable adaptive streaming in challenging mediums, such as cellular networks. State-of-the-art solutions adopt Deep Learning (DL) models to improve TP accuracy for various multimedia systems. This paper illustrates that designing blackbox TP engines that depend solely on the model’s capacity and power of learning does not achieve consistent accuracy across all throughput ranges. Additionally, we propose MATURE, a novel multi-stage DL-based TP model designed to capture network operating context to improve prediction accuracy. MATURE’s prediction involves characterising the operating context before estimating the network throughput. We show that MATURE delivers consistent, accurate prediction for all throughput ranges in both 4G and 5G networks. We also show that light-weight mature models that use quantized parameters maintain their accuracy while featuring up to 100x faster inference, thus making them suitable for mobile implementation. Our real video streaming experiments further show that MATURE improves the average user Quality of Experience (QoE) by up to 20% when compared to other throughput prediction methods.