How to use the Random Forest Visualizer
- 1. Pick dataChoose moons, circles, XOR or overlapping clusters, the number of points and a seed.
- 2. Set the forestChoose the number of trees, max depth, features per split and bootstrap sampling.
- 3. Grow itSee the vote map, the first six trees and the accuracy figures.
What a random forest does
A single decision tree splits the plane into rectangles and can overfit badly. A random forest grows many trees, each on a bootstrap sample (rows drawn with replacement) and choosing from a random subset of features at each split, then lets them vote. The trees make different mistakes, so the vote is smoother and usually more accurate than any one tree. Rows left out of a tree's bootstrap sample give a free test set, the out-of-bag (OOB) accuracy.
Worked example
With the defaults (two moons, 200 points, seed 7, 25 trees, max depth 5, 1 random feature per split, bootstrap on), the forest scores 92.0% on training data and 76.8% on 400 new test points, with 74.5% out-of-bag accuracy. The first tree on its own scores 69.8% on the same test set, so the forest improves on a single tree by 7 points.
Notes
- Trees are CART trees grown with Gini impurity, as in scikit-learn. With all features and no bootstrap, a tree we grow makes the same predictions as scikit-learn's DecisionTreeClassifier across a 40 × 40 grid.
- Deeper trees fit training data better; the forest's averaging limits the overfitting.
- OOB accuracy needs bootstrap sampling turned on.
- Try the XOR data with max depth 1: every tree is a single split, so the forest can't learn the four corners however many trees you add.
Frequently asked questions
How many trees do I need?
Accuracy usually stops improving after a few dozen to a few hundred trees. More trees don't overfit; they just cost more time.
Why only two features?
So the decision regions can be drawn on a flat chart. Real forests use as many features as your data has.
Is the random forest visualizer free?
Yes. You can open the random forest visualizer and enter everything without signing up. When you grow a forest for the first time, we ask once for your name and WhatsApp number; after that it, and every other Uncodemy tool, opens straight away.