Time for quality ML from event-based data
summary
Claims that computing features directly on raw event streams with a feature engine lets teams iterate far faster and avoid temporal leakage, using time-aware joins instead of brittle ETL. Example: a gaming customer produced 8 models in 2.5 weeks that were production ready, with feature development enabling 1000x more iterations and end-to-end delivery 20–26x faster from idea to scale. Advises against rebuilding low-level platforms and recommends off‑the‑shelf time-based feature tooling for behavioral ML.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:38A true story
- 3:04Eventbased data
- 4:48Data architectures
- 6:39What really happens
- 7:04Temporal leakage
- 9:42A new phase
- 10:51My point of view
- 11:19Feature engine
- 16:41How to get results
- 19:08How Kaskada can help
- 20:37Conclusion
related talks
40mCreating and operating ML models from event based data using feature stores
30mHow Feature Stores Enable Operational ML
26mReducing the cost of production AI
45mAnatomy of a production ML feature engineering platform
01h 28mFeature Engineering Made Simple
40mOvercoming data infrastructure limitations with a new paradigm for ML
20mSemblance: Feature Generation In Real-Time and Batch Without Time-Travel
45mBuilding a Feature Platform to Scale Machine Learning