Architecture

There are three common components to learning problems: the data to learn from, the learner that creates a mathematical model from the data, and the evaluation where problem-specific metrics provide insights on the performance of a given model. These three components are available in scikit-multiflow and allow 1) to easily setup and run experiments 2) easily extend existing methods for research purposes.

The BaseSKMObject class is the base class in scikit-multiflow. It is based on sklearn.BaseEstimator in order to support inter-framework compatibility. It adds extra functionality relevant in the context of scikit-multiflow.

Any stream model in scikit-multiflow is created by extending the BaseSKMObject class and the corresponding task-specific mixin(s), such as: ClassifierMixin, RegressorMixin, MetaEstimatorMixin and MultiOutputMixin

The ClassifierMixin defines the following methods:

  • fit – Trains a model in a batch fashion. Works as a an interface to batch methods that implement a fit() functions such as scikit-learn methods.

  • partial_fit – Incrementally trains a stream model.

  • predict – Predicts the target’s value in supervised learning methods.

  • predict_proba – Calculates the probability of a sample pertaining to a given class in classification problems.

The RegressorMixin defines the same methods for the regression setting with minor differences in the methods’ signatures.

The following UML diagram provides an overview of the base classes in scikit-multiflow with their methods.

base classes with methods

A stream model interacts with two other objects: a Stream object and (optionally) a StreamEvaluator object. The Stream object provides a continuous flow of data on request. The StreamEvaluator performs multiple tasks: query the stream for data, train and test the model on the incoming data and continuously tracks the model’s performance.

Following, is the sequence to train a stream model and track performance in scikit-multiflow using the Prequential evaluator.

prequential evaluation sequence

Note: The StreamModel class described in the original paper of scikit-multiflow has been replaced in version 0.3.0. The same functionality is now achieved by the above described combination of BaseSKMObject + the corresponding mixin.