Magic Mirror: Understanding the Diversified Network Conditions of Twitch Community
summary
Claims an unsupervised pipeline that turns client-side QoS time series into a global dictionary of playback behavior via hierarchical clustering and dynamic time warping, yielding 13 representative playback behaviors. The process maps sessions to clusters, computes network-parameter sets that reproduce each cluster’s playback metrics, and feeds those parameters into transcoder and player development. A tiny cluster C12 proves surprising by causing >10% of total buffering, and the method already speeds ABR rollout for low-latency mobile viewers.
AI-written summary · about
chapters (7)
- 0:00Start
- 8:53Room for Improvement (but how?)
- 9:41Challenge of Optimizing User Experience
- 11:00Long Cycle of Deployment and A/B Test
- 13:02Project Overview Step 2 Playback Behavior Clustering Analysis and Optimization
- 14:08Build the global "Dictionary" of Playback Behaviors (but it's expensive)
- 24:04Deep Dive on C12 group
related talks
22mProviding better video experiences for the next billion users.
17mMetrics Driven Obsession with Viewer Experience
14mAnalyzing Video Metrics like Richard Feynman
21mLearning in situ: a randomized experiment in video streaming
33mThe Challenges of Testing Video Playback at Scale
20mReconstructing proprietary video streaming algorithms
19mLow Latency Live from a Different Vantage Point
30mMLDS 2019: Anubhav Shrivastava of Voot at Machine Learning Developers Summit 2019