Decision tree builder

CART splitsGini / entropy / MSENumeric + categorical

Data

Sample dataset

Loading the sample data.

Target and settings

Target column–

Text targets are classified. A numeric target with more than 12 distinct values becomes a regression.

Split criterionGini

Both measure how mixed the classes in a node are.

Max depth4

Min split

Min gain

Min split is the fewest rows a node needs before it may split. Min gain is the smallest impurity drop worth a split. The tree rebuilds as you change a setting.

Inspect a row

Row number–

Or click any row in the data preview. The trace lists every decision the row meets on its way to a leaf.

How the tree is grown

Greedy CART-style splitting. At each node the builder tries every feature, keeps the single split with the largest impurity drop (the gain), then repeats on both halves. It never revisits an earlier choice.

Two kinds of question. Numeric features are cut at midpoints between neighbouring values. Categorical features test one category against all the others.

Impurity. Gini or entropy for classification, mean squared error for regression. A pure node scores 0.

Stopping. A node becomes a leaf at max depth, below min split rows, when the best gain is under min gain, or when it is already pure.

Missing values. Empty cells count as missing. A row missing the split feature is left out of both branches while training, and follows the no branch when traced.

Decision tree

Drag to pan. Ctrl + scroll or pinch to zoom.Drag to pan. Pinch to zoom.

Split
Leaf
Link width: rows on the branch
Traced row

Node colour: impurity

0 (pure)most mixed

Depth–

Nodes–

Leaves–

Training accuracy–

Building the sample tree.

Data preview

0 rows0 columns

Click a row to trace its path through the tree.