Inferring Video Streaming Quality of Experience at Scale using Incremental Statistics from CDN Logs.

With video streaming applications, end-user Quality of Experience (QoE) is generally estimated using metrics provided by the video players. While QoE monitoring is critical for Content Delivery Networks (CDNs), player metrics are not always provided due to technical, contractual or regulatory limitations. We propose to leverage CDN access logs to infer five common QoE player metrics, with the help of machine learning models. Our approach is based on constant-memory statistic accumulators, allowing large-scale analysis of video streams. We evaluated our implementation with more than 100 000 concurrent streaming sessions on a single CPU core, showing good correlation with QoE ground truth (ρ > 0.7, R2 > 0.5).