Assigned leaf node in decision_path()

[lorebon]

Hi Isak,

I have a question about the RandomShapeletForest method "decision_path": I was wondering if it's possible to modify the boolean 2D matrix in order to keep track of the assigned leaf node for each sample.

Is there another way to recover such information? Suppose that there are at least two leaves with the same class.
Thank you in advance!

[isaksamsten]

Hi!

Thanks for your question!

You can get the leaf node index using the apply method of ShapeletTreeClassifier

For example:

import numpy as np
from wildboar.ensemble import ShapeletForestClassifier
from wildboar.datasets import load_gun_point

X, y = load_gun_point()
f = ShapeletForestClassifier()
f.fit(X, y)

# Get the index of the node where each sample ends up
# eg leaves[0] contains the leaf node index for the first sample
leaves = f.estimators_[0].apply(X)

# count which leaf has the most samples in it
np.unique(leaves, return_counts=True)

You can also request it for the forest:

leaves = f.apply(X)
leaves[0, 0] # the index of the leaf of the first sample in the first tree
leaves[:, 0] # the index of the leaves for all samples in the first tree

While writing this, I realise this might not be what you are asking for? Are you asking for information about the majority class in each leaf?

You can get the probability distribution over labels in a tree from the tree_.value property:

f.estimators_[0].value

You'll note that some of these values are nan, inf etc, those values are for brach nodes, so to only get the values for leaf nodes we have to figure out which node indicies refer to leafs. We can do that using the tree_.left (or tree_.right) and find the values that are -1 which indicates that a specific index refers to a leaf

tree = f.estimators_[0]
leaf_proba = tree.tree_.value[tree.tree_.left == -1]

This will give us a matrix of (leaf_node, n_labels) with the distribution of labels that ended up in the leaf during training.

To get the class name that a leaf will assign we take the labels for the index of the maximal column from the value array:

leaf_nodes = tree.tree_.left == -1
labels = f.classes_.take(np.argmax(tree.tree_.value[leaf_nodes], axis=1))

We can use zip to get the node_index and the assigned label:

list(zip(np.nonzero(leaf_nodes)[0], labels))

This would return something like:

[(3, 2.0),
 (5, 2.0),
 (7, 1.0),
 (8, 2.0),
 (10, 1.0),
 (12, 2.0),
 (17, 2.0),
 (19, 2.0),
 (20, 1.0),
 (21, 2.0),
 (22, 1.0),
 (25, 1.0),
 (27, 2.0),
 (30, 1.0),
 (31, 2.0),
 (32, 2.0),
 (34, 1.0),
 (37, 1.0),
 (38, 2.0),
 (39, 2.0),
 (41, 2.0),
 (43, 1.0),
 (44, 2.0)]

Hope this helps!

I will close the issue (since its not an issue :))

We can continue the discussion under "Discussions"

Cheers,
Isak

[lorebon]

Thank you very much!