[논문 리뷰] River: machine learning for streaming data in Python

Jacob Montiel, Max Halford외 9명|arXiv (Cornell University)|2020. 12. 08.

Machine Learning and Data Classification참고 문헌 3인용 수 159

한 줄 요약

River는 Creme과 scikit-multiflow를 통합한 스트리밍/온라인 머신러닝용 Python 라이브러리로, 다양한 작업에 대한 통합 아키텍처, 파이프라인 및 벤치마크를 제공합니다.

ABSTRACT

River is a machine learning library for dynamic data streams and continual learning. It provides multiple state-of-the-art learning methods, data generators/transformers, performance metrics and evaluators for different stream learning problems. It is the result from the merger of the two most popular packages for stream learning in Python: Creme and scikit-multiflow. River introduces a revamped architecture based on the lessons learnt from the seminal packages. River's ambition is to be the go-to library for doing machine learning on streaming data. Additionally, this open source package brings under the same umbrella a large community of practitioners and researchers. The source code is available at https://github.com/online-ml/river.

연구 동기 및 목표

• Enable machine learning on streaming data with continual learning capabilities.
• Provide a flexible, unified architecture that supports multiple learning tasks (classification, regression, clustering, forecasting, anomaly detection).
• Offer transformers, pipelines, and performance evaluators to facilitate reproducible stream experiments.
• Benchmark River against existing streaming libraries to evaluate accuracy and speed.
• Promote accessibility and community adoption by consolidating tools under one open-source package.

제안 방법

• Architectural design with task-specific mixins to ensure compatibility across learning tasks (classification, regression, clustering, etc.).
• Core learning interfaces learn_one and predict_one (and related methods) for instance-incremental learning; supports batch-incremental via learn_many/predict_many.
• Pipelines to chain transformers and estimators, enabling preprocessing (e.g., StandardScaler) before learners.
• Efficient dictionary-based data containers with Cython-backed operations for fast feature handling and feature evolution.
• Instance-incremental and limited batch-incremental learning approaches to handle streaming data in real time.
• Benchmarking against existing libraries (GNB, LR, Hoeffding Tree) on Elec2 dataset, comparing accuracy and processing time.

실험 결과

연구 질문

• RQ1 How does River's accuracy compare to scikit-learn, Creme, and scikit-multiflow on standard streaming benchmarks like Elec2?
• RQ2 How do River's learn_one/predict_one and learn_many/predict_many performance compare in terms of speed and scalability across models?
• RQ3 Does River provide comparable or superior speed while maintaining competitive accuracy across key streaming tasks (classification, regression, clustering, forecasting, anomaly detection)?
• RQ4 What architectural features (mixins, pipelines, dictionary-based data containers) most effectively support model extensibility and feature evolution in streaming contexts?

주요 결과

• River achieves comparable accuracy to other streaming libraries on the Elec2 benchmark for Gaussian Naive Bayes, Logistic Regression, and Hoeffding Tree.
• River generally offers faster learn and predict times than competing libraries on the Elec2 dataset, indicating strong runtime performance.
• On the Elec2 benchmark, River's accuracy is closely aligned with scikit-learn, Creme, and scikit-multiflow across the evaluated models, demonstrating competitive predictive performance.
• The architecture emphasizes flexibility, ease of use, and a unified interface for both instance-incremental and limited batch-incremental learning, supporting diverse streaming scenarios.
• Benchmark results illustrate River’s capacity to serve as a go-to library for streaming machine learning with a broad set of supported tasks and efficient performance.