101 A/B Tests for personalized News
summary
Regiolab builds a central personalization platform for Omroep Gelderland using Primed.IO on Google Cloud Platform with Kubernetes, Spark, and Python. A/B headline tests and collaborative filtering run online; a Gaussian time decay (σ=3 days) on scores proves best so far. experiments report ~13,000 extra article reads in March and project a potential ~200,000 reads/month if scaled. Engineering uses precompute-and-cache (store top-20 recommendations in Redis) to hit ~100 ms p99 and handle 14,000 requests/min peaks.
AI-written summary · about
chapters (5)
- 0:00Regional news personalization intro
- 2:49Three A/B test pilots overview
- 8:06Engineering data science loop
- 12:44Scalable model delivery architecture
- 22:05Summary and future
related talks
25mBetter search relevance using Learning to Rank at mobile.de
03mServing Billions of Personalized News Feeds with AI
37mFood is personal! Design and impact of personalized recommendations in e-grocery
39mMachine Learning for Relevance and Serendipity
45mReal-time, Streaming Advanced Analytics....
42mA/B testing with real-world data to evaluate ML model updates
39mBuilding a personalised home feed for Germany’s biggest classifieds website
13mAlgorithmic Transparency in Personalized Advertising: Learnings from Building "Why Am I