[scikit-learn] Gradient Boosting: Feature Importances do not sum to 1

[Piotr Bialecki]

Hi Doug,

I modified your code a little bit to calculate the feature_importances of every tree of the forest.
In my opinion these feature importances should also sum to 1.0.

Since I could not access each DecisionTreeRegressor of your GradientBoositngRegressor, I created a new
ExtraTreeRegressor.

This is a bit off topic, but does anyone have an idea, why
type(ExtraTreesRegressor().estimators_)
results in a list and
type(GradientBoostingRegressor().estimators_)
results in an np.array?

Anyway, here is the code:

import numpy as np
from sklearn import datasets
from sklearn.ensemble import GradientBoostingRegressor, ExtraTreesRegressor

boston = datasets.load_boston()
X, Y = (boston.data, boston.target)

n_estimators = 712
# Note: From 712 onwards, the feature importance sum is less than 1
params = {'n_estimators': n_estimators, 'max_depth': 6, 'learning_rate': 0.1}
clf = GradientBoostingRegressor(**params)
clf.fit(X, Y)

feature_importance_sum = np.sum(clf.feature_importances_)
print "At n_estimators = %i, feature importance sum = %.20f" % (n_estimators , feature_importance_sum)


n_estimators_forest = 100
clf_forest = ExtraTreesRegressor(n_estimators=n_estimators_forest)
clf_forest.fit(X, Y)

feature_importance_sum_forest = np.sum(clf_forest.feature_importances_)
forest_feat_imp = [np.sum(tree.feature_importances_) for tree in clf_forest.estimators_]
print "At n_estimators = %i, feature importance sum = %.20f" % (n_estimators_forest, feature_importance_sum_forest)
for idx, imp in enumerate(forest_feat_imp):
    print "imp for tree %i: %.20f" % (idx, imp)


I suppose in each tree there is a small rounding error, summing up to the overall error.
So is this a bug or an inevitable rounding issue?


Greets,
Piotr

On 09.09.2016 03:51, Douglas Chan wrote:
Hello everyone,

I’d like to bring this up again to see if people have any thoughts on it.

If you also think this is a bug, then we can track it and get it fixed.  Please share your opinions.

Thank you,
-Doug


From: Douglas Chan<mailto:douglas.chan at ieee.org>
Sent: Wednesday, August 31, 2016 4:52 PM
To: Scikit-learn user and developer mailing list<mailto:scikit-learn at python.org> ; Raphael C<mailto:drraph at gmail.com>
Subject: Re: [scikit-learn] Gradient Boosting: Feature Importances do not sum to 1

Thanks for your reply, Raphael.

Here’s some code using the Boston dataset to reproduce this.

=== START CODE ===
import numpy as np
from sklearn import datasets
from sklearn.ensemble import GradientBoostingRegressor

boston = datasets.load_boston()
X, Y = (boston.data, boston.target)

n_estimators = 712
# Note: From 712 onwards, the feature importance sum is less than 1

params = {'n_estimators': n_estimators, 'max_depth': 6, 'learning_rate': 0.1}
clf = GradientBoostingRegressor(**params)
clf.fit(X, Y)

feature_importance_sum = np.sum(clf.feature_importances_)
print "At n_estimators = %i, feature importance sum = %f" % (n_estimators , feature_importance_sum)

=== END CODE ===

If we deem this to be an error, I can open a bug to track it.  Please share your thoughts on it.

Thank you,
-Doug


From: Raphael C<mailto:drraph at gmail.com>
Sent: Tuesday, August 30, 2016 11:28 PM
To: Scikit-learn user and developer mailing list<mailto:scikit-learn at python.org>
Subject: Re: [scikit-learn] Gradient Boosting: Feature Importances do not sum to 1

Can you provide a reproducible example?
Raphael

On Wednesday, August 31, 2016, Douglas Chan <<mailto:douglas.chan at ieee.org>douglas.chan at ieee.org<mailto:douglas.chan at ieee.org>> wrote:
Hello everyone,

I notice conditions when Feature Importance values do not add up to 1 in ensemble tree methods, like Gradient Boosting Trees or AdaBoost Trees.  I wonder if there’s a bug in the code.

This error occurs when the ensemble has a large number of estimators.  The exact conditions depend variously.  For example, the error shows up sooner with a smaller amount of training samples.  Or, if the depth of the tree is large.

When this error appears, the predicted value seems to have converged.  But it’s unclear if the error is causing the predicted value not to change with more estimators.  In fact, the feature importance sum goes lower and lower with more estimators thereafter.

I wonder if we’re hitting some floating point calculation error.

Looking forward to hear your thoughts on this.

Thank you!
-Doug


________________________________
_______________________________________________
scikit-learn mailing list
scikit-learn at python.org<mailto:scikit-learn at python.org>
https://mail.python.org/mailman/listinfo/scikit-learn



_______________________________________________
scikit-learn mailing list
scikit-learn at python.org<mailto:scikit-learn at python.org>
https://mail.python.org/mailman/listinfo/scikit-learn


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/scikit-learn/attachments/20160909/be9b6c01/attachment-0001.html>