[Extending proposal] MetaStream and BLAST

[maffei2443]

Hello, River Team.

I've developed a little library which implements two (or three) frameworks for algorithm selection, MetaStream and BLAST.
I would like to adapt the code, as well documenting it, so that it could be integrated into River.
Is that possible? Which patterns and "best practices" am I supposed to follow?

I know the current code needs a lot of improvement for people other than me using it of course, but I think it may fit the library purpose.

[MaxHalford]

Hello 👋

You linked the statsmodels library, but I'm guessing you referring to this repo of yours?

Anyway, we're definitely open to contributions :). Here is some advice:

• Start with a single method. You can always port the other one(s) later.
• CONTRIBUTING.md has instructions on our development workflow. Please let us know if it's not clear enough :)
• Reading some River code is the best way to understand how we do things.
• In your case, the methods you suggest are ensemble methods. Therefore I advise you get familiar with the ensemble module.
• The reason why they're ensemble methods is that all the base models are updated, and the output are getting aggregated in some way. This is different to the model_selection module, where the outputs are not aggregated, and instead a subset of models is updated and the best model is used to make predictions. Of course ensemble and model selection methods can be compared in terms of performance.

I hope that helps! We're all happy to answer questions you may have.