A growing number of video streaming networks are incorporating machine learning (ML) applications. The growth of video streaming services places enormous pressure on network and video content providers who need to proactively maintain high levels of video quality. ML has been applied to predict the quality of video streams. Quality of delivery (QoD) measurements, which capture the end-to-end performances of network services, have been leveraged in video quality prediction. The drive for end-to-end encryption, for privacy and digital rights management, has brought about a lack of visibility for operators who desire insights from video quality metrics. In response, numerous solutions have been proposed to tackle the challenge of video quality prediction from QoD-derived metrics. This survey provides a review of studies that focus on ML techniques for predicting the QoD metrics in video streaming services. In the context of video quality measurements, we focus on QoD metrics, which are not tied to a particular type of video streaming service. Unlike previous reviews in the area, this contribution considers papers published between 2016 and 2021. Approaches for predicting QoD for video are grouped under the following headings: (1) video quality prediction under QoD impairments, (2) prediction of video quality from encrypted video streaming traffic, (3) predicting the video quality in HAS applications, (4) predicting the video quality in SDN applications, (5) predicting the video quality in wireless settings, and (6) predicting the video quality in WebRTC applications. Throughout the survey, some research challenges and directions in this area are discussed, including (1) machine learning over deep learning; (2) adaptive deep learning for improved video delivery; (3) computational cost and interpretability; (4) self-healing networks and failure recovery. The survey findings reveal that traditional ML algorithms are the most widely adopted models for solving video quality prediction problems. This family of algorithms has a lot of potential because they are well understood, easy to deploy, and have lower computational requirements than deep learning techniques.
Unsupervised deep learning for real-time assessment of video streaming services
