Performance analysis of collaborative real-time video quality of service prediction with machine learning algorithms

With the exponential rise in the development, deployment and use of Internet applications in recent years, communication paradigms are continuously flooded with unmatched dimensions of real-time video traffic which amounted to 65.93% of the total worldwide traffic in January 2023 as reported by Sandvine. This paper presents an end-to-end framework where a series of machine learning and deep learning algorithms are used to accurately forecast the behaviour of network traffic parameters and video quality of service in a typical live video streaming environment. The ecosystem, developed with Java, is deployed on desktop applications, with the purpose of monitoring network activity while simultaneously generating a perceptual video mean opinion score through the blind image quality assessment technique via distortion aggravation. On the same wavelength, the architecture contains a centralised server application, responsible for receiving and channelling these functionalities from different desktop applications connected on the same network. These applications are equipped with time-series forecasting algorithms including multiple linear regression, multilayer perceptron and long short-term memory. With the training data collected in a real networking scenario, multiple sources of data are used to form a collaborative machine learning architecture. In total, 18 individual machine learning models were evaluated in terms of percentage accuracy and time complexity after optimising their hyperparameters. The collaborative multilayer perceptron regression algorithm built in this paper achieved a peak video quality of service prediction accuracy of 93.74% at small window sizes and proved to be efficient against non-collaborative methods across multiple machine learning frameworks. However, in terms of time complexity, the non-collaborative multiple linear regression model was able to predict the score at an average of 1.63 s at small window sizes while maintaining an accuracy of 88.61%. This trade-off highlights the potential for higher prediction accuracy at the expense of longer training periods and the availability of larger datasets. Moreover, across all tested scenarios, the long short-term memory algorithm exhibited the least favourable performance coupled with longer training times. The collaborative datasets used in this work showed the greatest significant improvement over non-collaborative ones for the multivariate models built with the inclusion of both network and video quality metrics, resulting in a peak increase of 2.70% in model accuracy for multilayer perceptron at large window sizes.