[MRG] Gradient boosting OOB improvement

[pprett]

This PR adresses issue #1802 by @yanirs.

It sets oob_score_ deprecated and introduces oob_improvement_ which gives the relative improvement of adding the i-th tree on the out-of-bag examples. The PR includes an example that shows how oob_improvement_ can be used to estimate the "optimal" number of iterations (basically an alternative to cross-validation).

[amueller]

Is this what the R package does? It makes sense to me.
Tests are missing, though ;)

[pprett]

yes (but R's gbm also smooths the OOB improvements using LOESS)

2013/7/22 Andreas Mueller notifications@github.com

Is this what the R package does? It makes sense to me.
Tests are missing, though ;)

—
Reply to this email directly or view it on GitHubhttps://github.com//pull/2188#issuecomment-21358376
.

Peter Prettenhofer

[agramfort]

pl is the standard :)

[amueller]

I like the example but I feel it is a bit hard to read. Could you add some comments please?

[arjoly]

n => n_samples?

[pprett]

@arjoly addressed - thx

[pprett]

@amueller I hope the example is clearer now

[arjoly]

(nitpicking) maybe import directly GradientBoostingClassifier

[arjoly]

I would add a test to check that oob_improvement_ is related to _oob_score_.
Furthermore you can think about adding a regression test.

Otherwise LGTM.

[pprett]

@arjoly I've added two regression tests and updated the narrative docs

[pprett]

@amueller @arjoly I think I've adressed the points you raised - it would be great if you could check - I'll then merge if its ok for you

[amueller]

Thanks for addressing the comments. 👍 for merge.

[arjoly]

LGTM!

[pprett]

merged - thanks for the reviews!