Accelerating Machine Learning Models for Video Streaming Traffic
Bence Ladóczki · 2025
The increasing popularity of video streaming platforms poses new challenges to network operators as a large proportion of the streaming traffic traverses the network in an encrypted format. Network administrators face significant challenges in accurately monitoring and assessing the Quality of Service (QoS) and the end users’ Quality of Experience (QoE). Tackling these difficulties, machine learning(ML)-based prediction models are frequently deployed to classify and recognise video streaming patterns and to gauge QoS/QoE. Here we follow the lines of scientific endeavours of other researchers and utilise ML models to infer user-side metrics relying solely on transport layer statistics. The performance of ML classifiers including a neural network is evaluated for YouTube Live, YouTube Gaming, Twitch and CNBC videos with compressed transport data. The ML models are trained on several gigabytes of data and for some parameters, more than 80% of accuracy is attained with a highly compressed dataset. Data compression is achieved by calculating the singular value decomposition of the dataset and keeping only the most significant vectors for prediction. The results are supplemented with comprehensive QoE metrics and the classifiers are evaluated against these metrics as well.