Using feedback loops to tune predictive models in a video ad marketplace
summary
Combines predictive models with PID controllers and cascading feedback loops to tune decision thresholds using historical completion rates plus real‑time player signals and logistic regression. Claims nearly twofold increase in available targeted ad slots while tightening campaign ad‑completion error (previously ~10–15% over) down toward zero by adjusting an adjusted performance goal in batch intervals. Built as part of the JW Player Marketplace; deployed across many campaigns and showing large reach gains with tighter fulfillment of advertiser targets.
AI-written summary · about
chapters (10)
- 0:00JW Player and ad marketplace
- 1:30Video ad auctions and targeting
- 5:43Predictive models for ad completion
- 10:17Feedback loops for model tuning
- 12:28Feedback loop demonstration
- 13:53Impact on campaign performance
- 15:37Challenges in ad marketplace
- 18:27Cascading feedback loops
- 21:54Conclusion and benefits
- 22:57Q&A
related talks
21mRise of the Machines: Removing the Human-in-the-Loop
27mLarge Scale Response Prediction to Serve Ads on LinkedIn
15mAuto-Adaptive Systems: Changing the Game for Marketplace Management
29mKnowledge Graph–Driven Feedback Loops to Optimize TikTok Shop Video Marketing
33mMachine Learning to moderate ads
17mPredict, Prune, Play: Efficient Video Playback Optimization Under Device...
54mOn How Machine Learning and Auction Theory Power Facebook Advertising
40mNext Best Action: Maximizing Customer Value