CFA: A Practical Prediction System for Video QoE Optimization
Why we need to do video QoE prediction?
Given a video session, if we can predict the QoE of different deployment (CDN, bit-rate and other deployments), we can easily choose a good deployment for this video session.
Challenges of current video QoE prediction
- Cannot capture complex factors that affect quality.
- Slow updating.
Many existing solutions fail on one or both counts. For instance, solutions that use less complex models (e. g., linear regression, Naïve Bayes, or simple models based on last-mile connection) cannot capture complex factors. Solutions that use complex model (e. g., SVM) take several hours to train the model.
So, in this paper, authors present CFA (Critical Feature Analytics). CFA (1) can capture complex factors. (2) updates quickly.
CFA
Each session has a feature vector which represent its properties.
X=[ASN, city, connection_type, player, site, live_or_vod, content_name, CDN, bit_rate]
Each session has 4 performance metrics.
Y=[bufratio, avg_bit_rate, join_time, start_failure]
本文正文内容与之前看过的paper基本一致，就略了。
Inspiration from this paper
First add content name and bit-rate into feature vector X.
This paper only build a prediction model. The decision model is based on C3.
C3 is a decision system which can choose a proper bit-rate and a proper CDN for each client.