Machine Learning in an Exchange Environment
summary
Claims that active exploration materially improves exchange revenue prove mostly false: the machine learners curse biases wins toward overpredicted ads, but the optimal bid uplift for the value of learning scales like 1/K^2 so efficiency gains fall as 1/K^4, making measured gains tiny (≈0.1% or less, sometimes negative). Exchanges already provide substantial free exploration, so practical rules—explore when cheap, favor frequent features, explore negative coefficients, use short fixed exploration windows and a learning account to split social vs. publisher payments—work best.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:33Display advertising overview
- 3:55Engineering the auction
- 8:27The machine learner's curse
- 13:00The value of learning
- 17:15Learning and externalities
- 23:41Practical implementation
- 30:43Limits of active exploration
- 39:09Conclusion and Q&A
related talks
39mMachine Learning Engineering
21mDemystifying Machine Learning for Programmatic UA & Retargeting
15mAI-Driven Application Optimization
27mEconometrics for Learning Agents
41mMachine Learning Models in a Dynamic Environment
30mActive Learning and Human-in-the-Loop Machine Learning
29mAI and ML for Computational Advertising
19mUnveiling the Potential: Exploring Upselling, pLTV, and ML for Operations