Combining feature selection with meta learning

[Stephen Piccolo]

I have a somewhat broad question. I work with datasets that have thousands of features. I'm wondering whether anyone has done feature selection within the context of meta learning. Is that a thing? If so, could you point me in the direction of how that might work within the context of OpenML?

Thanks in advance,

-Steve

[Joaquin Vanschoren]

Hmm, not that I know of :).

--
You received this message because you are subscribed to the Google Groups "OpenML" group.
To unsubscribe from this group and stop receiving emails from it, send an email to openml+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/openml/3b4ec8cf-8ee7-42c5-b2c0-058b2edc2600n%40googlegroups.com.

[Joaquin Vanschoren]

Well, meta-learning AutoML systems that do pipeline construction (e.g. Autosklearn) do this implicitly.

E.g. they find similar datasets (based on meta-features) and then apply the preprocessing steps (including feature selection) that were used in the best pipelines on those datasets.

[Stephen Piccolo]

Thanks for your reply. Just to clarify...in this scenario, they won't necessarily use the same features that were identified in the best pipelines for those datasets, correct? They would use the same feature selection techniques (if they were used in those pipelines) to select features based on the current dataset. Is that correct?

[Joaquin Vanschoren]

Yes, exactly.

You could meta-learn which exact features to select if you have a scenario with many tasks with the same features. That’s very specific, and I don’t know of metalearning work in this direction. 

--

Joaquin Vanschoren, Ph.D., ir.

Associate Professor 

Department of Mathematics and Computer Science
Eindhoven University of Technology (TU/e)
https://joaquinvanschoren.github.io/