Data-Driven Quality of Experience Prediction Model and Measurement of Multimedia Streaming Services using Machine Learning

Abstract

The Internet traffic today is mostly multi-media traffic due to the exceptionally expand- ing interest in Over The Top (OTT) services like Facebook, YouTube, Netflix etc. It

has proven to be extremely difficult for Internet service providers (ISPs) to meet their customers’ needs in terms of Quality of Experience (QoE) because of the proliferation

of networking data, especially video streaming. Therefore, QoE modelling and measure- ment of multimedia services is an open research area for the research community. Due

to ever-increasing user demand for multimedia services, ISPs and OTT providers require

innovative solutions for QoE prediction of HTTP Adaptive Video-streaming (HAS) ap- plications as most of the video services over the internet are HAS-based. Therefore,

the QoE prediction model will lead towards identifying the root causes for QoE im- pairments and understanding the impact of different Key Quality Indicators (KQIs).

The primary objective of this study is to propose supervised-learning-based QoE pre- diction ensemble Voting Regression (VR) and Stacking Regression (SR) models based

on machine-learning algorithms such as Random Forests (RFs), Support Vector Re- gression (SVRs), Stochastic Gradient Descent (SGD) and Multilayer Perceptron (NN)

models considering appropriate QoE influencing factors. We utilize Waterloo Streaming

Quality-of-Experience Database for more accurate prediction of QoE over the multi- media video streaming services in this study. This work has a multi-fold contribution:

First, the data set was optimized using four feature selection techniques based on ma- chine learning also including Principal Component Analysis (PCA) to investigate the

impact of different KQIs and retain the most appropriate ones in the feature-engineering stage. Secondly, making k-fold validations and hyper-parameter tuning of standalone ML models was adopted to check the accuracy of each model over the given data set

in the model optimization and training stage. Thirdly, upon these hyper-parameter- tuned base ML models, ensemble VR and SR models were created. In the final stage,different ML models were evaluated based on learning curves, execution times, training times and performance metrics for comparative analysis among various features obtained from different feature selection techniques and then analyzed the algorithm which suits best for estimated QoE prediction. The results show significantly higher scores of R2 i.e 0.852367 and a significant improvement of 4.64% in terms of R2 was observed as compared to previous studies. Finally, the lower values of MAE, MSE and RMSE i.e 0.085513, 0.220756 and 0.469846 were obtained respectively, while showing the highest PLCC value of 0.92539 and SRCC value of 0.875782 while predicting QoE.