Ewa Dominowska, Engineering Manager, Facebook @ MLconf SEA
summary
Facebook describes building an ads ML stack that selects from over a billion ads for over a billion people, splitting the problem into information retrieval, click and conversion prediction models, and a pricing auction pipeline; prototypes in R and one‑day datasets fail at scale so engineering must productionize online models with continuous counts; experimentation requires careful traffic splits, seasonality checks and accounting for AB testing pitfalls and model interactions; centralized applied ML teams outperform single embedded hires.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 0:33Agenda
- 0:59Building a team
- 7:29Selecting a team leader
- 9:19Where to put the team
- 11:05Motivation
- 14:07Flowchart
- 17:23Problem Approach
- 19:05Evaluation
- 22:27Practical Lessons
- 25:00Summary
related talks
54mOn How Machine Learning and Auction Theory Power Facebook Advertising
03mApplied machine learning at Facebook
19mRushin Shah, Engineering Manager, Facebook
33mMachine Learning to moderate ads
23mThe unreasonable effectiveness of data for better monetisation
27mML Platform Journey
21mDemystifying Machine Learning for Programmatic UA & Retargeting
29mBuilding ML Platform Capabilities at Global Scale