Moving from Offline to Online Machine Learning with River
summary
River claims to make online ML practical in Python by learning one observation at a time and reusing scikit‑learn–style APIs for classification, regression, anomaly detection and time‑series forecasting. It provides a metrics package and progressive validation for delayed labels, copes with concept drift, and reduces memory/CPU compared with batch retrains. Real deployments require a message broker (Kafka) and a stream processor; examples show replaying event logs to bootstrap models and using Quick streams / Quickclad for pipelines. Sliding windows and many models are implemented and demonstrated working.
AI-written summary · about
chapters (23)
- 0:00Introduction
- 2:07Talk overview
- 3:54Databases vs message brokers
- 4:56Streaming data history
- 6:39Python dominance
- 7:43Quickstream and Formula 1 background
- 9:04Quickstream features
- 10:50Flux capacitor analogy
- 11:11Offline ML definitions
- 14:03Data journey real-time to at-rest
- 15:27Online vs offline inference
- 16:29Online learning with streams
- 17:53Benefits of online ML
- 19:18Concept drift resilience
- 21:23Online ML use cases
- 21:45Batch vs stream complementarity
- 23:10Kappa architecture
- 24:35River library origin
- 24:56River features
- 26:21Dictionary data handling
- 27:02Metrics in River
- 28:05River documentation and resources
- 28:48Closing remarks
related talks
23mMoving from Offline to Online Machine Learning with River
31mMoving from Offline to Online Machine Learning with River
30mReconceptualizing Machine Learning for the Real-time World
41mStreams to the River: Powering an Online XGBoost Classifier with Kafka
22mOnline Machine Learning with Creme
37mAutomated Machine Learning: Building a Pipeline
36mActive-Active Machine Learning Infrastructures For Training And Prediction
40mReal-time Learning with Spark