DecisionTreeClassifier max_leaf_nodes tree.tree_.value is set to 0 for non leaf nodes. Normal behaviour?

[robdempsey]

Hi

>>> sklearn.__version__
'0.15.2'

If I leave out max_leaf_nodes or set it to None when accessing tree.tree_.value all non leaf nodes values are set to zero and leaf node is correct. If I set the value to something like 10 the values are present?

dtree = tree.DecisionTreeClassifier(criterion = "entropy", min_samples_leaf = 1000, max_depth=4, max_leaf_nodes = None)

values    = tree.tree_.value

Iterating through a branch

[[ 0.  0.]]
[[ 0.  0.]]
[[ 0.  0.]]
[[ 1767.  6014.]]
18

dtree = tree.DecisionTreeClassifier(criterion = "entropy", min_samples_leaf = 1000, max_depth=4, max_leaf_nodes = 10)

values    = tree.tree_.value

Iterating through a branch

[[  45455.  166666.]]
[[  41958.  142857.]]
[[  8392.  28572.]]
18

Rob

[amueller]

I think that is because different splitting stragegies are used depending on whether you set max_leaf_nodes or not. Some of them don't store the internal counts, as they are not needed for prediction.

[glouppe]

The reason was only to speed things up.

@arjoly I wonder however if this was indeed needed at all. It seems that these counts should be there already (since we need them to compute the impurity decrease), hence storing them shouldnt hurt much. What do you think?

[arjoly]

If benchmarks show that the tree growing procedure is not significantly slower, I am +1.
Note that there are more efficient memory-wise alternatives to store the node values for multi-output decision tree.

since we need them to compute the impurity decrease

At the moment, we don't need those. Could you clarify?

[robdempsey]

They are a useful output if you want to present the results to a non-technical audience and don’t want to explain entropy :-)

[amueller]

well you could reconstruct them from the output you have, there is a 1:1 mapping between leaves and path ;)

[trevorstephens]

#3735 ... just sayin... ;-)

[andosa]

PR created #4655

[glouppe]

This is now fixed :)