Random Forest Visualization - Interactive Ensemble Learning

Dataset

Forest hyperparameters

Number of trees 9

Max depth 4

Feature sampling per split 1 of 2

Forest performance

Training accuracy

–

Out-of-bag accuracy

–

Unanimous votes

–

Avg. tree depth used

–

How it works

Each of the trees ringing the plot is grown from its own bootstrap sample: N points drawn with replacement from the dataset, so roughly 63% of points appear at least once and the rest are left "out of bag" for that tree. At every node the tree only gets to look at a random subset of features (x, y, or both) — this is what decorrelates the trees so the ensemble beats any single one of them. The best split is the one that most reduces Gini impurity:

Gini(S) = 1 − Σ pᵢ²
Gain = Gini(parent) − [ (|L|/|S|)·Gini(L) + (|R|/|S|)·Gini(R) ]

The tree keeps splitting on whichever (feature, threshold) pair maximizes that gain until it hits max depth or a pure node. The forest floor shows the ensemble's answer: every cell is colored by majority vote across all trees. Out-of-bag accuracy scores each tree only on the points it never trained on — a free validation set baked into bagging.

Number of trees — more trees average away individual mistakes; OOB accuracy stabilizes as this grows.
Max depth — deeper trees fit tighter boundaries per tree but overfit their own bootstrap sample more.
Feature sampling — "1 of 2" forces each split to gamble on x or y alone (true random-forest behavior); "2 of 2" degrades toward plain bagging, and trees become more correlated.

Real-world relevance: this bootstrap + feature-subsample + majority-vote recipe is exactly what scikit-learn's RandomForestClassifier runs at production scale, just with hundreds of trees and higher-dimensional feature vectors.

Class 0 Class 1 Tree leans class 1

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.