ML: Predictive analytics in Ad Engineering
summary
Company runs 400,000 active communities with ~300 million monthly unique users and ~10 terabits in the warehouse. For full-traffic rollouts when A/B is infeasible, they use interrupted time series, fit models with Facebook Prophet, generate 1,000 posterior predictive samples and trim 5% tails to report a median percent lift (about 10%) and an uncertainty interval. For daily health they use sroka to pull data from Ad Manager, Athena and partners (10+ bidders) and mostly rely on rule-based anomaly detection because simple thresholds catch the majority of issues quickly.
AI-written summary · about
chapters (6)
- 0:00Introduction to Fandom
- 4:50Ad Engineering Team Overview
- 9:29Key Ad Metrics Explained
- 15:43Predictive Analytics Use Cases
- 24:47Anomaly Detection Approach
- 34:46Takeaways and Q&A
related talks
11mData science: Next-gen performance analytics
18mData-driven Decision Making in Ad-Tech
30mMMM and MTA: The Future of AI-Driven Marketing
53mReal-Time Big Data Analytics with R
18mBuilding Models Using Ad Sales as an Example
40mApplying Data Science and Analytics in Marketing
46mFrom correlation to causation: measuring ad effectiveness at scale
15mThe Secret Life of Data - Part 2