AI and ML for Computational Advertising
summary
Computational advertising frames each impression as an optimization: choose ad a to maximize value b, solving roughly 100 billion auctions/day in under 100 ms. Systems use DSPs and ad exchanges, real-time servers reading predictive models and feature lookups, with response prediction, time-series forecasting and mechanism design feeding the bid. Typical scale metrics cited: 400,000 requests/sec, 100 TB/day of logs; observed sparsity about ≈0.1% CTR and ≈1% post-click conversion. Practical problems are latency, few positive labels, multi-party allocation and privacy-aware targeting.
AI-written summary · about
chapters (9)
- 0:00Intro
- 1:55Optimization problem
- 6:30Advertising
- 10:40Conversion Funnel
- 14:33Optimization Problems
- 18:50How do all these systems work
- 21:25Forecasting
- 23:50Marketing Hypothesis
- 27:05How Does This Scale
related talks
27mAI and analytics for better marketing performance
33mAdTech
54mOn How Machine Learning and Auction Theory Power Facebook Advertising
44mOptimizing Marketing Spend with Auto-Bidding Algorithms
33mMachine Learning to moderate ads
24mAI-Assisted Media Planning
23mThe unreasonable effectiveness of data for better monetisation
21mQuestions you Should be Asking Your DSP