Algorithmic Transparency in Personalized Advertising: Learnings from Building "Why Am I
summary
Meta rolls out version 3.0 globally in February 2023 and shows concrete engineering choices: a distilled explainable model built from the complex multi-stage ads ranker, a feature‑value translator that maps embedding IDs to topic sentences, and a 1–2s latency budget. Evaluation uses an AD flip rate test plus normalized entropy, R2 and human relevance judgments; user helpfulness surveys increase, but metrics drift when production models change, so the system works but requires continuous retraining and operational investment.
AI-written summary · about
chapters (7)
- 0:00Algorithmic transparency in ads
- 1:12ML transparency for ads
- 2:58Evolution of transparency features
- 6:28Engineering explainable models
- 10:40User-friendly explanations
- 11:42Measuring explanation quality
- 13:08Q&A
related talks
22mSession: Meta Ads: The Next Chapter
19mMachine Learning for Privacy-Preserving Ads
22mCommunicating Algorithmic Process in Online Behavioral Advertising
21mDemystifying Machine Learning for Programmatic UA & Retargeting
21mRise of the Machines: Removing the Human-in-the-Loop
30mMMM and MTA: The Future of AI-Driven Marketing
18mExtract Immense Value from Creative Optimisation
21mGrowth Hacking 2022