tree.export_graphviz() returns wrong count of examples in leaf nodes when using example weighting

[dragoljub]

When using the DecisionTreeClassifier() sample_weight parameter, and weighting examples by class lable such as 2:1 for class A versus class B, the nvalues in the leaves of the tree produced by tree.export_graphviz() shows a duplicate count for the number of examples for the weighted class.

This is misleading because the sum of each classes nvalues should equal the total number of examples in the parent node. It seems that there is a bug where the feature weighting factor is not removed when exporting the decision tree.

Here is the classifier with weighting:

clf = tree.DecisionTreeClassifier(
    criterion='gini', splitter='best', max_leaf_nodes=10,
    min_samples_split=10, min_samples_leaf=2, max_features=None,
    random_state=None).fit(X, y, sample_weight=df.LABEL.map({'A':2, 'B':1})) 

Here is the produced tree.dot file with the duplicate values shown in bold:

digraph Tree {
0 [label="X[106] <= 0.0203\ngini = 0.0731494237327\nsamples = 119377", shape="box"] ;
1 [label="X[34] <= 1266.6650\ngini = 0.0661567536396\nsamples = 118307", shape="box"] ;
0 -> 1 ;
3 [label="gini = 0.0578\nsamples = 111941\nvalue = [ 210926. 6478.]", shape="box"] ;
1 -> 3 ;
4 [label="gini = 0.2103\nsamples = 6366\nvalue = [ 10016. 1358.]", shape="box"] ;
1 -> 4 ;
2 [label="X[83] <= 6728.4551\ngini = 0.386307949063\nsamples = 1070", shape="box"] ;
0 -> 2 ;
5 [label="gini = 0.4994\nsamples = 364\nvalue = [ 254. 237.]", shape="box"] ;
2 -> 5 ;
6 [label="gini = 0.1669\nsamples = 706\nvalue = [ 68. 672.]", shape="box"] ;
2 -> 6 ;
}

Here is what the tree looks like with duplicate leaf example counts for the weighted class:

[trevorstephens]

Within clf.tree_.value the samples are weighted and this is where export_graphviz grabs that part of the node text. There is no other attribute within the tree object to get unweighted node sample counts by class, so I don't think this can be easily changed without modifying the way the actual trees are built and what is stored in the object.

Adding a separate parameter to export_graphviz such as class_weights might be a way to unravel the values, but that would assume constant weights by class, which might not always be the way it is used.

#3735 is adding new functionality where the samples and values can be represented as proportions, rather than raw counts which could help with interpretability in this case, but the output will still represent the weighted proportion, not raw sample counts.

[amueller]

I would favor adjusting the docs to reflect what is happening (which I did in #3828). What do you think?

[dragoljub]

Its good to have this note in the docs. I would also clarify that in the leaf node it says "values" where it shows class sample size. Is it possible to show something like samples = [Class A = 1234, Class B = 456]?

[amueller]

Maybe values is not the most intuitive label. Your proposal would be very easy to implement. Opinions @glouppe @arjoly @larsmans @pprett ?
Changing the dot output might break things downstream I guess, but I think we emphasized the dot file was for visualization purposes only.

[galthram]

You can change the output using this function I made. Maybe there is a better way, but it works:

#Here Y_true are the true classes the tree is trying to predict. Y_true is a pandas Series. clf is
the tree classifier.
dot_data = tree.export_graphviz(clf, out_file=None,
                                 filled=True, rounded=True,
                                 special_characters=True
                                 )
def ajust_tree (dot_data, Y_true):
    n_samples= len(Y_true.index  )
    n_classes= len(  Y_true.unique() )
    class_w = n_samples / (n_classes * Y_true.value_counts() )
    s=dot_data
    inicio = 0
    i=0
    
    while (i > -1):
        i = s.find("value = ", inicio)
        if (i > -1):
            j = s[ (i + 8) :].find("]")
            u=s[i+9: i +8 + j ].split(", ")
            valor = np.array( u  ).astype(float) 
            valor_n=(valor / class_w).round().astype(int)
            nova_string=", ".join(str(x) for x in valor_n)
            s=s[:i+9] + nova_string + s[ i +8 + j:]
            inicio = i +1
    return s