river

Why we included this project

Most machine learning assumes you have the whole dataset in front of you, but River is for the cases where data arrives as a continuous stream. Instead of retraining a batch pipeline on a schedule, you feed observations one at a time into its incremental classifiers, regressors, clustering, and anomaly-detection algorithms, updating the model as each event arrives. Concept-drift detection is built in, so the model can notice when the data distribution shifts and adapt rather than quietly degrading. That makes it a natural fit for real-time analytics, fraud and intrusion detection, IoT sensor feeds, and other streaming workloads where training a fresh model each time is too slow or wasteful. The scikit-learn-style API keeps integration straightforward, and the thorough documentation and published JMLR paper give teams a solid basis for committing to online learning in production.