Thomas J. Fan

> Would this feature be available soon anytime on the master ? When there is enough interest and time for other core members to review it. This PR was updated...

If we are looking strictly at citations, it looks like most papers cite Mark Hall's thesis with [3383 citations](https://scholar.google.com/scholar?hl=en&as_sdt=0%2C33&q=Correlation+Based+Method+feature+selection&btnG=). Extending from this work is a paper by Lei Yu and...

I think it's still worth doing. Looking over the comments, I think the issue was around adjusting the `plot_permutation_importance_multicollinear.py` example and mixing together concepts of feature importance and feature selection....

Currently, `SVC` ignores parameters depending on the kernel. For invalid combinations, I agree with @NicolasHug that we should error instead. I think we should error when the parameter is **not**...

I would prefer to deprecated first to be on the safe side. For the first PR, let's do this for `gamma` first to see how other reviewers feel about it....

The ignoring behavior is [explicitly documented](https://scikit-learn.org/stable/modules/generated/sklearn.svm.SVC.html#sklearn.svm.SVC) which makes me think it was intentional. I do want to get to raising an error at some point and would be +1 on...

> Why not just go through the usual deprecation cycle? This is more user friendly than a breaking change. Looking at this again, if we were to change behavior, I...

Using `'auto'` to mean `1/n_feature` is kind of weird in itself. To move forward we can also rename `'auto'` to ` reciprocal_n_features` ?

A workaround is to start with an empty dict, iterate through the tasks, and add the results to the dict in the dynamic workflow: ```python from dataclasses import dataclass from...

Sorry, I misunderstood the issue. My workaround does not resolve your issue with having to use another task to construct the dataclass.