Random pings to people at the sprint @ogrisel @betatim

Still have to look at the example...

Repository health increased by 0.00% when pulling 9fb131f on glouppe:tree-apply into 6e54079 on scikit-learn:master.

• 1 new problem was found (including 0 errors and 1 code smell).
• No problems were fixed.

I improved the example a bit, but I am not very happy with it at all in fact. I would vote for removing it. What do you think?

Note that we already have an example showing the capabilities of trees for embedding.
http://scikit-learn.org/dev/auto_examples/manifold/plot_lle_digits.html#example-manifold-plot-lle-digits-py

Coverage increased (+0.0%) to 95.12% when pulling 9d4bb37 on glouppe:tree-apply into 6e54079 on scikit-learn:master.

I was looking at this yesterday and thought of rebasing it :) Thanks for picking it up. I'll have a look later.

Do you have the picture genreated by the example?

This is the plot that the example produces, it doesn't say much to me.

I agree it's not the most convincing example. I think the "sparse" feature extraction point is not well made by using a tree with only two leaves. As for part 1 of the example, I am not convinced why I would want to transform the classification problem that I care about into one where I'd just predict the noisy leaves of a decision tree trained on the original problem. The only reason I can find is to do model compression, but why would I compress a tree? Maybe a whole forest.

How about an example that focuses on the second part, but actually creates a Transformer class and pipelines it into a linear classifier, would that make sense?

Come to think of it, given the digits embedding example referenced by @glouppe, the only usefulness of a tree.apply example would be to show users how to roll their own. Might be simpler to just remove the example completely.

+1 on removing the example.
Possible other example: comparing RandomTreeEmbedding, RandomForestClassifier and GradientBoostingClassifier for feature extraction and compare PCA vs a trained classifier on them?

The GradientBoostingClassifier doesn't have apply (as it is not necessarily tree-based) so we could use the tree apply.

Possible other example: comparing RandomTreeEmbedding, RandomForestClassifier and GradientBoostingClassifier for feature extraction and compare PCA vs a trained classifier on them?

The GradientBoostingClassifier doesn't have apply (as it is not necessarily tree-based) so we could use the tree apply.

Ah, this would be a great example indeed. GB not being necessarily tree-based is a great reason to provide this as an example rather than as a transformer in the library. In the old PR, @ogrisel pointed to a notebook where he used this:

http://nbviewer.ipython.org/github/ogrisel/notebooks/blob/master/sklearn_demos/Income%20classification.ipynb

with a link to this paper:

Practical Lessons from Predicting Clicks on Ads at Facebook Junfeng Pan, He Xinran, Ou Jin, Tianbing XU, Bo Liu, Tao Xu, Yanxin Shi, Antoine Atallah, Ralf Herbrich, Stuart Bowers, Joaquin Quiñonero Candela International Workshop on Data Mining for Online Advertising (ADKDD)

https://www.facebook.com/publications/329190253909587/

hehe yeah Ralf Herbrich was my boss at Amazon, which is why I thought about it ;)

Sorry for the delay. I removed the example, improved tests and now raise NotFittedError instead of ValueError.

LGTM. whatsnew?

LGTM. whatsnew?

Done :)

Opened #4549 in case anyone is bored ;)

Coverage increased (+0.0%) to 95.16% when pulling 6ea24c7 on glouppe:tree-apply into b1eaac0 on scikit-learn:master.

A last review? Should be quick, there is not much. CC: @vene

I improved the documentation. Merging.