Using Performance Signals to Improve Video Recommendations
summary
Client devices send bandwidth estimation and real‑time playback signals (starts, stalls, wait time, memory/CPU hints) to ranking models that downrank risky videos or promote lighter content; the server returns predicted watch time and other cues to drive a video prefetch strategy that downloads fewer chunks for short predicted views and more for long ones. Models use deep nets with video, user and social embeddings plus watchtime percentiles, yielding fewer stalls, higher resolution, lower data waste and measurable engagement gains.
AI-written summary · about
chapters (8)
- 0:00Performance signals intro
- 1:50Recommendations with device signals
- 2:55Bandwidth estimation for ranking
- 4:48Stall signals for feeds
- 6:14Watch time prediction strategies
- 8:00Prefetching impact on experience
- 10:28Bidirectional signal exchange
- 12:14Future of personalized video
related talks
17mPredict, Prune, Play: Efficient Video Playback Optimization Under Device...
14mAnalyzing Video Metrics like Richard Feynman
15mDashlet: Taming Swipe Uncertainty for Robust Short Video Streaming
26mBandwidth Ballet: Optimizing Video Performance for Web Maestros
22mProviding better video experiences for the next billion users.
41mDelivering Fast Video without the Buffering
26mOptimizing Video Performance for Web Developers
13mSENSEI: Aligning Video Streaming Quality with Dynamic User Sensitivity