Large Scale Response Prediction to Serve Ads on LinkedIn
Talkat Parallel and Distributed Algorithms for Inference and Optimization (Simons Institute, Oct. 21–24, 2013) · 2013-10-21..10-24·
summary
LinkedIn describes a production system that predicts ad click probability by combining a cold-start logistic regression trained on billions of observations and thousands of features with a warm-start per-ad random effects model fit in parallel per ad, served via MapReduce. Training scales with a consensus ADMM scatter/gather routine (initialization and adaptive step cut wall-clock training from ~10 to ~2 hours). Serving uses Thompson sampling jittering on per-ad posteriors for explore/exploit; online tests show consistent eCPM and prediction improvements.
AI-written summary · about
chapters (8)
- 0:00Overview of LinkedIn ads
- 1:14The advertising system flow
- 3:29Automated format selection
- 8:22Predicting click-through rates
- 12:36Cold and warm start models
- 15:27ADMM for model training
- 18:27Thompson sampling results
- 23:39Summary and future outlook
related talks
33mMachine Learning to moderate ads
53mReal-Time Big Data Analytics with R
30mPredictive Analytics in Social Media and Online Display Advertising
41mChallenges of Machine Learning at Scale
54mOn How Machine Learning and Auction Theory Power Facebook Advertising
31mLarge Scale CTR Prediction Lessons Learned
46mFrom correlation to causation: measuring ad effectiveness at scale
29mAI and ML for Computational Advertising