Thanks for the PR.
I am not that familiar with the algorithm, and it would be great if you could add examples that compare this against elasticnet and linear discriminant analysis to show the benefit.

Also, is this seems to be the same as
make_pipeline(PCA(), LogisticRegression(), LogisticRegression()), right?
If that is the case, I don't think adding an extra model is warranted, maybe an example?

It is not that easy to do for LogisticRegression, but it will be with #4242.

Hello,

It is similar but not equivalent. Supervised PCA also contains an extra step of filtering out useless attributes. So, the steps are

1. Filter out attributes: Fit a model having only a single feature as input, and if the coefficient is above a threshold, then keep it.
2. Conduct PCA on the reduced dataset
3. Fit the final model.

The first step is something that can require a few lines of code, so having a new model makes life a bit easier :)

In the book "The Elements of Statistical Learning" there were some examples on where this technique would be better against elastic net. I am not sure if they contained examples that include LDA. I can look into it.

Best regards,
Stelios

Sorry, I misread the code then.
So it is equivalent to make_pipeline(LogisticRegression(), PCA(), LogisticRegression())?
The first logistic regression will drop features below a threshold, then you do a PCA and then train a model on the outcome. I'm not sure what the PCA is for, though? I'll have a look at ESL....

Ah, ok, you do drop the number of components. And they use univariate selection. So it is make_pipeline(SelectKBest(), PCA(), LogisticRegression()).
They do use a more fancy univariate selection criterion in the first step, which is more appropriate for survival analysis. That gives them an advantage on their data.

I don't think this particular pipeline deserves it's own estimator.

There's already a somewhat similar example on scikit-learn's documentation. It's only missing the PCA step.
http://scikit-learn.org/stable/auto_examples/feature_selection/feature_selection_pipeline.html

closing as no reply and no excitement. This is a pretty straight-forward pipeline imho.