Machine learning at scale enables Communications Service Providers (CSPs) to deliver better customer experience.
We live in an encrypted world. This is indispensable from privacy and security perspectives. However, encryption produces significant technical challenges for CSPs to assess the quality of services they provide. And in no other application, these limitations are more evident than in video streaming.
CSPs quantify the Quality of Experience (QoE) of video streaming with network Quality of Service (QoS), application QoS, and user QoE. However, the application QoS and the user QoE are often lost due to encryption, leaving only the network QoS indicators available. CSPs nevertheless must determine the user and application QoE in order to validate their network planning and optimisation initiatives and ensure consistent delivery of expected QoE.
“CSPs must deploy machine learning techniques that can infer, based on what is visible in network packet flows, the de facto QoE of video sessions being consumed by their clients on the ground.”
Machine learning deployment in CSPs has to cope with huge scales. In one example from Niometrics’ experience, edge nodes received in total 3.2 Petabytes of YouTube video volume per day, representing more than 400 million YouTube video sessions per day at a peak throughput of 714.5 Gbps. That is to say, more than 2 billion predictions per day were needed to indicate the de facto QoE of all Youtube sessions.
At Niometrics, we use our prorietary, fully-integrated technology stack to process high volumes of rich data streams. By adopting machine learning algorithms, we infer QoE from encrypted streams and enable CSPs to monitor and analyse the video streaming services delivered through their network.
Watch Constantinos Halevidis, VP of Data Science at Niometrics, explain how his team applied machine learning techniques at scale to extract KQIs such as initial buffering time, resolution, and stall events with a small hardware footprint.