Machine Learning Engineering
summary
Claims that machine learning is not magic and that long-lived systems succeed only when requirements, KPIs and data dependencies are explicit. Presents a concrete recommender with CTR threshold example and an ad-tech three-model RTB pricing system (forecast auctions, win probability, value) with a business baseline test. Describes evaluation and rollout tools: causal impact and multi-armed bandits for model replacement, adversarial drift detection and monitoring, and explains predictions with LIME RICE DOLLY.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 4:33Terminology
- 5:19Process
- 13:12Success Criteria
- 23:21Monitoring
- 23:55Data dependencies
- 26:00Data monitoring
- 27:46Challenges
- 28:23Testing
- 30:44Causal Analysis
- 32:21Multi unbounded problems
- 33:37Expect the unexpected
- 34:06Information
- 34:35Explanation
- 38:21Checklist
- 39:01Conclusion
related talks
48mPrinciples and Practices of Building Machine Learning Systems
21mDemystifying Machine Learning for Programmatic UA & Retargeting
49mThe Impact of Machine Learning in Companies
18mReliable Maintenance of Machine Learning Models
07mPredictive Modeling and Machine Learning
27mKeys to Building Machine Learning Systems
12mMachine Learning for Adaptive, Self-Optimizing Systems
44mSRE and ML: Why It Matters