Skip to content

Decision trees

fannar.tree interprets a fitted scikit-learn DecisionTreeClassifier. A tree has no activations and no gradient, so it gets its own instantiation of the sources, one "layer" per depth. At depth \(l\) each object has:

  • path: indicator of the nodes visited up to depth \(l\) (the state of the tree);
  • slack: \((x_f - t)/\sigma_f\) at every internal node already crossed, i.e. how far the object is from switching branch (the tree's sensitivity);
  • classes: the class distribution of the node the object reaches at depth \(l\).
from sklearn.tree import DecisionTreeClassifier
from fannar.tree import TreeInterpreter

tree = DecisionTreeClassifier(max_depth=5).fit(X_train, y_train)
interp = TreeInterpreter(tree, df_test, y_test, X_train=df_train)   # slack in training std units
report = interp.evaluate()
interp.explain_pair(0).top(5)
interp.explain_errors().top(5)

The joint pipeline is KernelPipeline(["path", "slack", "classes"]) and the baseline is the path alone.

argument meaning
X_train training rows. Their per-feature standard deviation is \(\sigma_f\) in the slack
scale an explicit per-feature \(\sigma_f\) (overrides X_train). Without either, \(\sigma_f = 1\)
depths the depths used as layers (default 1 … max_depth)
reference_model another tree (e.g. shallower or randomised) used as the floor

X can be a DataFrame (its columns become feature_names). The labels can be any of the tree's classes, strings included. They are mapped to indices into tree.classes_, which also become class_names, so the readings and explanations work for any label type.

fa.Interpreter(tree, X, y, scale=...) still works as in earlier versions.

Full example

examples/tree_wine.py fits a depth-4 tree on wine with text labels and a tree fitted on shuffled labels as the floor. See Examples.

Things to watch

  • X_train for the slack. Without it, the slack is in the raw units of each feature. A feature measured in thousands then dominates the distance to the thresholds. Pass the training rows, or scale=.
  • A floor for trees. There is no "untrained tree". A tree of the same depth fitted on shuffled labels is a useful floor: same structure, no signal.
  • Depth 1. The first layer has a single split. The path alone puts every object at one of two points, and only the slack spreads them. Expect the joint and path readings to differ little there, and more from depth 2 on.
  • Ensembles. Random forests and boosting are not single trees. Use fa.Interpreter with an extractor built from model.apply(X) (see Choosing a module).