Explainable classification rules
Need an interpretable baseline for labeled tabular data.
Scikit-Learn · Class
DecisionTreeClassifier
A supervised decision-tree classifier that learns feature-based rules for binary or multiclass prediction. Its split criterion, growth limits, weighting, randomness, and pruning behavior are configurable.
How it works
The classifier recursively chooses feature-based splits to form a tree, using a selectable impurity or information-gain criterion. Training stops according to depth, sample, leaf, impurity, or leaf-count constraints; the fitted tree then predicts classes and probabilities for new samples.
Syntax
class sklearn.tree.DecisionTreeClassifier(*, criterion='gini', splitter='best', max_depth=None, min_samples_split=2, min_samples_leaf=1, min_weight_fraction_leaf=0.0, max_features=None, random_state=None, max_leaf_nodes=None, min_impurity_decrease=0.0, class_weight=None, ccp_alpha=0.0, monotonic_cst=None)Parameters
| Name | Type | Description |
|---|---|---|
criterion | {"gini", "entropy", "log_loss"} | Split-quality function: gini, entropy, or log_loss; default gini. |
splitter | {"best", "random"} | Split-selection strategy: best or random; default best. |
max_depth | int or None | Maximum tree depth; None means no explicit depth limit. |
min_samples_split | int or float | Minimum samples required to split an internal node; integer count or fractional proportion. |
min_samples_leaf | int or float | Minimum samples required in a leaf; integer count or fractional proportion. |
min_weight_fraction_leaf | float | Minimum weighted fraction of the total input weight required in a leaf. |
max_features | int, float, {"sqrt", "log2"} or None | Number of features considered for a split, or None for all features. |
random_state | int, RandomState instance or None | Controls estimator randomness. |
max_leaf_nodes | int or None | Maximum number of leaf nodes; None means unlimited. |
min_impurity_decrease | float | Minimum weighted impurity decrease required for a split. |
class_weight | dict, list of dict or "balanced" | Class weights, or None for uniform weighting. |
ccp_alpha | non-negative float | Complexity-pruning parameter; larger values prune more aggressively. |
Returns
A fitted DecisionTreeClassifier instance from fit; predict returns class labels and predict_proba returns class probabilities.
Key points
- The class implements supervised classification with a tree of feature-based decisions.
- criterion supports gini, entropy, and log_loss for evaluating candidate splits.
- splitter chooses the split strategy independently at each node; best is deterministic in strategy, while random samples candidate splits.
- max_depth limits how many levels the learned tree can contain; None allows expansion until another stopping condition applies.
- min_samples_split prevents an internal node from splitting unless enough samples reach it.
- min_samples_leaf enforces a lower bound on samples represented by every leaf.
Common mistakes
- Defaults can create fully grown, unpruned trees; constrain complexity or apply pruning when overfitting is a concern.
Example: Fit and predict with a decision tree
from sklearn.tree import DecisionTreeClassifier
X = [[0, 0], [1, 1], [0, 1], [1, 0]]
y = [0, 1, 1, 0]
model = DecisionTreeClassifier(max_depth=2, random_state=0)
model.fit(X, y)
print(model.predict([[0, 0], [1, 1]]))
print(model.predict_proba([[0, 1]]))Creates a bounded-depth classifier, fits it to labeled samples, then predicts classes and class probabilities.
When to use it
Multiclass classification
Class labels include more than two categories.
Class-imbalance handling
Training data contains unequal class frequencies.
Regularized tree modeling
Model complexity must be constrained or pruned.
Interview questions
QWhat does DecisionTreeClassifier do?
It is a supervised tree model that learns feature-based decision rules for classification. It supports binary and multiclass classification and can return both predicted classes and class probabilities.
Follow-up: How would you reduce overfitting? Limit max_depth, increase min_samples_leaf, restrict max_leaf_nodes, or use cost-complexity pruning with ccp_alpha.
QHow do criterion and splitter affect training?
criterion selects the split-quality measure: gini uses Gini impurity, while entropy and log_loss use Shannon information gain. splitter controls whether the best split or a random candidate strategy is used.
Follow-up: When might random splitting help? It can add variation among trees or models, but a fixed random_state is needed for repeatable results.
QHow do sample-size constraints control the tree?
min_samples_split sets the minimum samples needed to split an internal node, while min_samples_leaf sets the minimum samples permitted in a leaf. Integer values specify counts; float values are fractions converted using a ceiling operation.
Follow-up: Why use min_samples_leaf? It prevents leaves supported by too few observations and can make predictions less sensitive to individual samples.
Related
DecisionTreeRegressorExtraTreeClassifierExtraTreesClassifierRandomForestClassifierHistGradientBoostingClassifiersklearn.tree.plot_tree