Trees & ensembles · /tree

Decision tree

CART, one split per Step click. Each split is the winner of an exhaustive gain scan (watch it in Internals); hovering a tree node highlights its rectangle on the plane.

class 0class 1filled = train, hollow = testhover to probe · click to pin

step 0impurity —train acc —test acc —

Data

datasetpoints150noise0.15seedsplitmouse

Gini impurity

I_{G} = 1 - k \sum p_{k}^{2} = 1 - p_{0}^{2} - p_{1}^{2} = I_{G}

Bound to the leaf under your cursor (or the last split node). Hover the plane to change it.

Entropy

H = - k \sum p_{k} lo g_{2} p_{k} = H

Information gain of the last split

Δ = I (parent) - \frac{n _{L}}{n} I (L) - \frac{n _{R}}{n} I (R) = Δ

The split chosen is the candidate with the highest Δ — watch the scan in Internals.

What's simplified here

The threshold scan is exhaustive over both features — with 2 features and small n this is the exact CART split, not an approximation.
No pruning; growth stops at max depth, min samples per leaf, or purity.