Skip to content

Extend a model tree

Start with a regression or classification pathway. Then choose the change that matches the modeling problem. These options are different decisions, not a checklist to enable together.

Need Technique Read next
Outcome is censored at a known bound Tobit leaves Tobit example
Domain limits on direction or combinations Monotone or interaction constraints Constraint example
Preserve segments while updating leaf models Leaf retraining Retraining examples
Control fit complexity and flexibility Growth and leaf settings Growth knob tour
Different outcome distributions Distributional families Distributional leaves
Predict only from approved input columns Feature and interaction controls Parameters

Censored outcomes: Tobit

Tobit is a model for a latent response observed only above a censoring boundary. Structural zeros, truncation and missing outcomes are different problems. Its interpretation depends on the assumed latent model and censoring mechanism.

from random import Random
from mltreelib import ModelTree

rng = Random(12)
x = [[rng.uniform(-2, 2)] for _ in range(160)]
y = [max(0.0, 0.5 + 2 * row[0] + rng.gauss(0, 1)) for row in x]
model = ModelTree(max_depth=0, min_rows=10, leaf="linear",
                  leaf_family="tobit", tobit_threshold=0.0)
model.fit(x[:120], y[:120], col_names=["signal"])
print("Latent predictions:", model.predict(x[120:])[:3])

The response above is the fitted latent expectation, so negative values are possible even though observations are censored at zero. Evaluate the desired observed outcome separately. Correct censoring assumptions do not guarantee an unbiased result on every dataset. See Tobin (1958).

Constrain shape and interactions

from mltreelib import ModelTree

x = [[float(i)] for i in range(40)]
y = [float(i // 10) for i in range(40)]
model = ModelTree(max_depth=2, min_rows=5, leaf="mean", monotone=[1])
model.fit(x, y, col_names=["volume"])
print("Predictions:", model.predict([[0.0], [10.0], [20.0], [30.0]]))

A positive constraint encodes a domain requirement. Check predictions on meaningful grids and held-out data. A check on the healthcare example found two decreases with linear leaves when age increased. The same check found no decreases with mean leaves. This is a tracked behavior gap; a finite grid does not prove monotonicity elsewhere. Do not assume local coefficient signs guarantee global monotonicity. Interaction constraints restrict paths and should also be checked with split_paths().

Retrain leaves with fixed partitions

from mltreelib import LeafControl, ModelTree

x = [[float(i)] for i in range(40)]
y = [float(i // 10) for i in range(40)]
model = ModelTree(max_depth=1, min_rows=5, leaf="mean")
model.fit(x, y, col_names=["volume"])
before = model.splits()
model.retrain(x, [v + 1 for v in y], control=LeafControl(leaf="mean"))
print("Splits preserved:", before == model.splits())
print("Updated prediction:", model.predict([[5.0]]))

This example mutates the model. Use appropriate independent data for evaluation and check that each leaf has sufficient new observations. Fixed partitions can become stale when the population changes. See the hierarchical retraining guide for alternative feature sets and routing.

Tune deliberately

Start with max_depth and min_rows to control partition size. Compare mean and linear leaves only when their assumptions and interpretation fit the outcome. Tune on a validation set and keep a final test set untouched. The knob tour is a parameter reference, not a recommended sequence or a substitute for evaluation.