online-ml/river
River
River est un framework Python pour le machine learning en ligne (online machine learning), conçu pour entraîner et évaluer des modèles sur des données en streaming. Il permet un apprentissage incrémental en mettant à jour les paramètres du modèle une observation à la fois, éliminant le besoin de stocker des jeux de données d'entraînement complets en mémoire.
La bibliothèque se distingue par un système dédié de détection de dérive de concept (concept drift) qui surveille les changements dans les distributions de données pour déclencher l'adaptation du modèle. Elle fournit également un framework de validation progressive qui simule un déploiement en temps réel en testant les modèles sur des échantillons avant de les utiliser pour l'entraînement.
Le système couvre un large éventail de capacités de streaming, incluant l'ingénierie de caractéristiques (feature engineering) en temps réel, la prévision de séries temporelles et la détection d'anomalies en ligne. Il prend en charge l'apprentissage non supervisé via le clustering incrémental et les arbres de décision, ainsi que l'agrégation ensembliste et les politiques de bandit pour la sélection de modèles.
Le projet inclut des utilitaires pour l'ingestion de données en streaming à partir de sources telles que des fichiers CSV et des API, ainsi que des outils pour calculer des statistiques courantes et des esquisses de données (data sketches) économes en mémoire.
Features
- Incremental Model Training - Implements incremental learning by updating model parameters one observation at a time to eliminate the need for storing full training datasets.
- Online Learning - Provides a framework for training and updating machine learning models on streaming data one observation at a time.
- Incremental Model Updating - Updates model weights one observation at a time to enable learning without storing full datasets in memory.
- Incremental Training - Fits models to data streams by processing observations one by one to eliminate the need for full dataset storage.
- Model Performance Evaluators - Performs progressive validation to test model performance on a stream by comparing predictions against ground truth labels.
- Model Pipelines - Merges multiple preprocessing and modeling steps into a single sequence to process data streams.
- Progressive Stream Validators - Simulates real-time deployment by testing each observation before using it for model training.
- Stochastic Gradient Optimizers - Updates model weights using stochastic optimization algorithms such as Adam, SGD, and RMSProp.
- Streaming Feature Engineering - Converts raw streaming data into meaningful representations using categorical encoding and interaction analysis.
- Time Series Forecasting - Predicts future values in sequential data streams by learning from historical observations in real time.
- Drift Detection - Identifies changes in the underlying distribution of a data stream to trigger model retraining or alerts.
- Input Distribution Drift - Includes a dedicated system for monitoring input distribution drift to trigger model adaptation or alerts.
- Streaming Preprocessing Pipelines - Chains preprocessing and estimation steps into sequential workflows for transforming raw streaming features.
- Distribution Monitoring - Detects concept drift by comparing data distributions between recent and historical sliding windows.
- Online Running Statistics - Offers utilities for calculating running statistics and memory-efficient data sketches under fixed memory constraints.
- Streaming Data Sketches - Maintains memory-efficient sketches of data streams to track unique elements, heavy hitters, and histograms.
- Progressive Model Evaluation - Tests model accuracy on a stream using progressive validation to simulate real-world deployment.
- Anomaly Detection - Provides algorithms and tools for identifying outliers or unusual patterns in data streams.
- Filter-Based Feature Selection - Implements filter-based feature selection using statistical tests and variance thresholds to improve model efficiency on streaming data.
- Streaming Bandit Selectors - Implements bandit policies to balance exploration and exploitation for selecting the best model or action in streaming scenarios.
- Imbalanced Stream Resampling - Adjusts class distributions of incoming streaming data through resampling to prevent bias in imbalanced datasets.
- Incremental Decision Tree Learners - Grows decision trees incrementally using the Hoeffding bound to ensure statistical significance of splits.
- Ensemble Learning - Implements ensemble methods like bagging and boosting to improve predictive performance and robustness on streaming data.
- Model Ensembling - Combines multiple incremental models through bagging, boosting, or stacking to improve overall predictive performance.
- Stochastic Gradient Descent - Optimizes model parameters using iterative updates based on the gradient of individual samples.
- Mini-Batch Processing - Handles small groups of observations using data frames to balance batch efficiency with online learning requirements.
- Anomaly Detection - Identifies unusual observations in live data streams by scoring samples based on evolving distributions.
- Feature Scaling and Transformations - Scales numeric values and encodes categories in real time to ensure data compatibility with algorithms.
- Streaming Data Sketches - Implements memory-efficient probabilistic structures to track statistics of high-volume data streams.
- Personalized Ranking Optimizers - Provides algorithms for maximizing the posterior probability of user preferences to sort items for recommendation tasks.
- Streaming Data Clustering - Groups incoming observations into clusters incrementally without storing the full dataset.
- General Machine Learning - Framework for general-purpose online machine learning.
- Incremental Learning - Online machine learning.
- Online Machine Learning - Library for online machine learning.
