Chapter 2 · Machine learning basics
Random forests and ensembles
Many varied models voting together beat one model on its own.
Chapter 2 · Machine learning basics
Many varied models voting together beat one model on its own.
One tree: jagged, because it memorised the odd fruits.
Each little map is one tree trained on a shuffled copy of the same fruit. The big map is their vote: darker colour means more trees agree.
A tree grown until every leaf is pure memorises odd fruits. Show it slightly different data and it draws a quite different map.
Green is ripe, coral is not ripe. Fruit with a wrong label are mixed in on purpose. Add trees and watch the forest score on 300 brand-new fruit.
One tree: its score depends on luck. Single trees here score between about 70% and 95%.
A simplified teaching version: forty made-up fruits, fifteen in a hundred with a wrong label, and 300 clean fruit as the test. Real forests also let each split pick from a random subset of features.
Why does each tree get its own shuffled copy of the data?
What does the forest do to reach its answer?
You add the 30th tree to a forest of 29. What usually happens to accuracy?