search

sktime / sktime

A unified framework for machine learning with time series
https://www.sktime.net
BSD 3-Clause "New" or "Revised" License
7.97k stars 1.38k forks source link

[ENH] estimator model categorization tags - e.g., "deep learning", "foundation model", "classical" #6669

Open fkiraly opened 5 months ago

fkiraly commented 5 months ago

Idea: should we have a tag or multiple that categorize models by algorithm kind?

What I mean is not the scientific type - forecaster, classifier - but the "fitting principle" that the model uses. High-level categories would be "classical statistical", "machine learning (excluding neural networks)", "deep learning (excluding foundation models)", "foundation models (= deep learning trained on large corpus)".

Similar categorizations exist in classification models already ("distance based", "deep learning"), and pairwise transformers also have a tag specifying whether they are a distance (triangle equality) or kernel (positive semi-definiteness).

This kind of information may be particularly helpful in retrieval and estimator selection for the user, e.g., by using the estimator search function in the overview page: https://www.sktime.net/en/stable/estimator_overview.html

Opening up the discussion, @sktime/core-developers

yarnabrina commented 5 months ago

We can have that tag. If I am honest, not sure if it will be widely used by users in programming part, but I agree that this can definitely help to look at options combined with tag based filters in updated docs pages.