Machine Learning for Relevance and Serendipity

on YouTube ↗39m

" summary

Prismatic builds a real-time personalization engine that combines NLP topic classification (about 5000 topics) with a logistic regression ranker and feature engineering to score user–article pairs in a few hundred milliseconds. Training uses saved session snapshots, offline prototyping and a per-user weight transformation to avoid heavy runtime work. The team fixes presentation bias, statistical bleeding and Simpson’s paradox by removing raw publisher features and careful evaluation; deployed results show roughly 50% more shares, 25% higher first‑item click rate, 15% longer sessions and 15% fewer “crap” sessions.

AI-written summary · about

" chapters (24)

1. 0:00Machine Learning for Relevance and
2. 1:06The Discovery Problem
3. 2:50How we do it
4. 3:54People who use the term 'secret sauce' know about neither
5. 4:44Talk Outline
6. 7:48Our Topic Classifier
7. 10:16Interest Scores
8. 10:58Saved Session Data
9. 13:06You are all unique snowflakes
10. 13:51How do we know if we've done a good job?
11. 16:02Formula for a Machine Learning Problem
12. 16:32Collect Data • Save session information (described earlier)
13. 17:22Featurize Data
14. 20:45Training the model (numeric optimization)
15. 21:22Classify New Data
16. 22:02Original formulation weight vector is universal, features are user-specific
17. 23:45The Development Cycle
18. 24:56Data bugs (are the worst)
19. 27:32Presentation bias
20. 28:21Ranking vs Normal ML
21. 31:10Statistical Bleeding
22. 33:45Simpsons Paradox
23. 37:24Summary • We collect information about our users' preferences and dispreferences
24. 38:18Prismatic Backend Team

" related talks

• 32mLeveraging recommender systems to personalise search resultsSoraya Hausl · PyData London 2017
• 27mTransform online shopping experiences with personalized recommendationsSarah Strobhar · AWS Machine Learning Summit 2021
• 41mRecommender Systems: Crafting Effective Personalised ExperiencesKim Falk · Øredev 2025
• 23mML-Powered Search & Recommendation 101Sergey Polyashov · Conf42 Machine Learning 2025
• 28mThe Power of Personalized Customer Experiences Through AI InnovationsN. Trajkov · DSC Europe 23
• 43mRecommendation Systems in Market ResearchDavit Abgaryan · PyData Yerevan 2022
• 29mThe Challenges of Building a Personalized Recommendation System in E-CommerceP. Poznic · Data Science Conference 4.0
• 03mServing Billions of Personalized News Feeds with AIMeihong Wang · Artificial Intelligence Conference 2018 New York