On How Machine Learning and Auction Theory Power Facebook Advertising

on YouTube ↗54m

" summary

Facebook predicts probabilities for dozens of user events (clicks, likes, hides, watch time) with logistic regressions and boosted trees, tests new features rapidly, and multiplies those probabilities by expressed per-event values to score items; the live system handles 2 million advertisers and must make trillions of ad predictions per day so it filters candidate ads, uses a dynamic controller to vary ranking depth, applies VCG allocation across organic, sponsored and ecosystem bids, and uses multiplicative pacing to smooth budgets.

AI-written summary · about

" chapters (11)

1. 0:00Start
2. 16:10The Ads Machine Learning Problem
3. 19:50Boosted decision tree + logistic regression
4. 20:49The Need for Speed
5. 24:57The Facebook Ads Auction
6. 31:47The VCG Auction
7. 34:03Anatomy of a VCG Auction
8. 38:39Unified Value Function
9. 41:21Budget Constraints
10. 43:25Solution 2: Discount bidding
11. 45:45Two-pronged Value Maximization

" related talks

• 30mPredictive Analytics in Social Media and Online Display AdvertisingMahesh Kumar · PASS Business Analytics Conference 2015
• 33mMachine Learning to moderate adsVaibhav Singh, Jaroslaw Szymczak · PyData Berlin 2017
• 22mGenerating a Billion Personal News FeedsEwa Dominowska · MLconf 2016 Seattle
• 29mAI and ML for Computational AdvertisingAbraham Bagherjeiran · Global Artificial Intelligence Conference 2020
• 26mEwa Dominowska, Engineering Manager, Facebook @ MLconf SEAEwa Dominowska · MLconf Seattle 2015
• 27mLarge Scale Response Prediction to Serve Ads on LinkedInDeepak Agarwal · Parallel and Distributed Algorithms for Inference and Optimization (Simons Institute, Oct. 21–24, 2013)
• 22mCommunicating Algorithmic Process in Online Behavioral AdvertisingMotahhare Eslami, Sneha R. Krishna Kumaran, Christian Sandvig, Karrie G. Karahalios · CHI 2018
• 23mThe unreasonable effectiveness of data for better monetisationTed Verani · Pocket Gamer Connects London 2018