scikit-multiflow allows to easily design and run experiments and to extend existing stream learning algorithms.[3] It features a collection of classification, regression, concept drift detection and anomaly detection algorithms. It also includes a set of data stream generators and evaluators. scikit-multiflow is designed to interoperate with Python's numerical and scientific libraries NumPy and SciPy and is compatible with Jupyter Notebooks.
Implementation
The scikit-multiflow library is implemented under the open research principles and is currently distributed under the BSD 3-clause license. scikit-multiflow is mainly written in Python, and some core elements are written in Cython for performance. scikit-multiflow integrates with other Python libraries such as Matplotlib for plotting, scikit-learn for incremental learning methods[4] compatible with the stream learning setting, Pandas for data manipulation, Numpy and SciPy.
Components
The scikit-multiflow is composed of the following sub-packages:
anomaly_detection: anomaly detection methods.
data: data stream methods including methods for batch-to-stream conversion and generators.
drift_detection: methods for concept drift detection.
evaluation: evaluation methods for stream learning.
lazy: methods in which generalisation of the training data is delayed until a query is received, i.e., neighbours-based methods such as kNN.
meta: meta learning (also known as ensemble) methods.
trees: tree-based methods, e.g. Hoeffding trees which are a type of decision tree for data streams.
History
scikit-multiflow started as a collaboration between researchers at Télécom Paris (Institut Polytechnique de Paris[5]) and École Polytechnique. Development is currently carried by the University of Waikato, Télécom Paris, École Polytechnique and the open research community.
