Sequential Learning for Modeling Video Quality of Delivery Metrics

Abstract

Video streaming traffic growth poses a challenge for many video content providers to maintain high video quality on their networks. Modern day networks are highly dynamic due to the adaptation of routing protocols, time-varying loads, and adaptive codecs. We consider how these dynamics should be incorporated in learning algorithms. We ask how can sequential learning be used in order to account for time-varying trends in learning algorithms that perform forecasts in time-varying video delivery systems? We propose two approaches. The first is called ASAPjitter. It uses a Recurrent Network (RN) to forecast gradients in a time series of jitter measurements for a video stream and a Convolutional Neural Network to classify the forecasts. The second approach is called FEATjitter. It maps the jitter time series to a higher dimensional feature domain. This novel feature domain transform has the aim of enhancing the quality of prediction and classification of future video jitter measurements. Jitter captures varying congestion levels present in the network. Three parameters are extracted from the jitter time series: its period, base and rate of decay. These parameters are used to train an artificial RN to perform forecasts. A batch learning approach such as the Neural Network (NN) or otherwise referred to a Multi-Layer Perceptron (MLP) is used for classifying the feature domain forecasts. Experimental results demonstrate how sequential learning can be used effectively to predict and classify jitter using Deep Learning (DL) frameworks. The accuracy achieved by ASAPjitter is 84.5 % compared to 91.6% for FEATjitter. Performance gains in terms of increased classification accuracy are due to the learning algorithm operating in this feature-space. Performance gains in terms of forecasts are more effective in the time-domain. Our approach contributes to the development of more effective algorithms for managing video streaming traffic in dynamic network environments.