Predicting the QoE in Adaptive Video Streaming Using Weighted Averaging Federated Learning

Abstract

Recently, assessing the QoE in adaptive video streaming systems has become an interesting area of research since it directly gauges customer satisfaction. Quality of Experience (QoE) models potentially suffer from data availability issues and a lack in preserving sensitive data of clients. The ITU-T standards provided a guideline for low-cost practical objective QoE assessment to obtain the streaming, non-streaming parameters, and their corresponding accurate MOS scores. The existing QoE prediction literature came with a flavour of implementing the models separately or sharing the QoE data between distinct devices. This work simulating the user interaction with the online video distributers, and obtaining the labelled QoE data using the ITU-T P.1203. In addition, it proposes the Weighted Averaging Federated Learning (WAFL), an enhanced federated learning implementation using the feed forward neural networks to predict the QoE. The WAFL preserves the recent privacy requirements by avoiding sharing the entire data among the distributed models. The training is implemented in a sequential manner amongst the collaborated nodes and enhances the global model by aggregating the shared learned weights through the distributed learning rounds. The achieved QoE prediction accuracy is compared to the isolated learning and the traditional federated learning. The proposed QoE prediction provided an accuracy of 86.68% in estimating the QoE using a small number of streaming and non-streaming QoE parameters.