[Feature Proposal] Integration of the StDE (Self-Tuning Drift Ensemble) for Concept Drift Detection

[gysakurai]

Category: Feature Proposal / Idea
Paper: Paper Link

Proposal: Integrating the StDE Tool

I am proposing the integration of a new tool, StDE (for Self-Tuning Drift Ensemble), into the River ecosystem. This tool, recently developed and detailed in my paper, aims to significantly enhance the project's capabilities in the area of Concept Drift Detection.

Why StDE?

The StDE tool addresses a goal in the current implementation by providing:

1. [Key Benefit 1]: StDE adapts to data stream changes by dynamically adjusting the number and configuration of its base learners based on voting behavior.
2. [Key Benefit 2]: Already utilizes the River Framework, with its base detectors such as a DDM and EDDM.
3. [Key Benefit 3]: Tests on homogeneous and heterogeneous drifts demonstrated StDE’s robustness and efficiency in maintaining a lightweight structure.

Proposed Implementation/Integration

I believe the StDE tool could be integrated into the project in the following way:

• As a New Module: The StDE tool could be introduced as a new detector option within the existing library structure, leveraging the River Framework already in use. This would allow the community to readily explore and validate StDE alongside existing base detectors like DDM and EDDM across various data types and drift scenarios.
• Dependencies: : StDE already utilizes the River Framework, including its base detectors . Therefore, there are minimal new external dependencies required for its integration, simplifying maintenance and installation.

Supporting Documentation

The full technical details, methodology, and evaluation results are available in my paper:

• Title: A Self-Tuning ensemble approach for drift detection
• Link: Paper Link

I am eager to hear feedback from the maintainers and the community to determine if StDE genuinely adds value to the River toolset. If the consensus is positive, I'd like to discuss the most suitable approach for integration.

Specifically, I would appreciate input on:

Value Proposition: Does StDE's approach align with River's current direction and offer a significant enhancement over existing detectors?

Integration Path: What is the best technical approach for integrating StDE (e.g., as a direct component within the main library, or an official extension/contrib module)?

I am fully available to answer any questions and respond to all comments. I am ready to develop and submit a Pull Request (PR) once a clear implementation path is agreed upon.

Thank you for your time and consideration!

[MaxHalford]

Hello there @gysakurai. We do not think it is a good idea to integrate new methods that have just been published. This is very much akin to how sklearn decides what to integrate in their library or not.

[gysakurai]

Thanks for the response, got it! I appreciate the feedback. Since @smastelini has been following the project, I thought it would be a good idea to integrate this method.

[smastelini]

I believe your drift detection ensemble could live in the river-extra repo (it is a bit dusty but has the purpose of showcasing algorithms that may some day come to the main repo)

[gysakurai]

That makes perfect sense! Thanks so much, @MaxHalford and @smastelini for feedbacks.