Moving Forward Through The Darkness
summary
States specific modeling blind spots and how they are fixed: data–problem mismatch that makes post logs useless for purchase prediction, loss of signal during vectorization, label drift from new items, and biased objective functions. Shows a deployed news recommender built from user view click and user non-click signals, combines weak rankers by weighted sum and reinforcement feedback, uses SGD, Kafka/streaming pipelines and metric learning for better similarity; reports production use and improved recommendations.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:22Model brightness in data science
- 9:07Real problems to ML solutions
- 12:00Modeling procedure
- 13:05News recommendation case study
- 23:59Bias sources in modeling
- 27:44Data generation relevance
- 31:41Vector conversion pitfalls
- 34:14Framework bias in ML
- 36:23Handling new labels
- 39:17Conclusion
related talks
29mFurther enhancements of recommender systems using deep learning
01h 21mUnlocking Personalization: A Deep Dive into Modern Recommendation Algorithms
58mRecommender Systems – A Walkthrough
01h 31mHands on - Build a Recommender system
23mML-Powered Search & Recommendation 101
32mLet's talk about Recommendation Systems!
38mRecommendations at Rappi: from MVP to Data Product
21mDemystifying Machine Learning for Programmatic UA & Retargeting