What does a leaf contain?

The prediction for its path: an average or majority computed from the training cases that landed there, plus their counts.

What is a minimum leaf size?

A rule refusing splits that would create leaves below a set population, protecting against tiny-sample predictions.

How do leaves relate to overfitting?

Growing many small leaves lets the tree memorize training cases; limiting leaf size and pruning are the brakes.

How many leaves should a tree have?

As many as validation justifies: enough to capture real structure, few enough that each carries a respectable population, with pruning making the trade explicit.

AI Glossary / AI and machine learning / Beginner

Leaf

Definition

A leaf is an end node of a decision tree: the terminal box where a case finally lands, carrying the tree's prediction for every case that arrives there.

What it means

A decision tree works by repeated splitting: each internal node asks a question, each branch takes an answer, and the walk ends at a leaf, which holds the prediction for all cases that followed that exact path.

The leaf is where the model's knowledge physically lives: a churn rate, an average price, a majority class, computed from the training cases that landed in that box. Everything upstream is just routing.

Leaf size is a health metric: leaves built on thousands of cases give stable predictions, while a leaf holding eleven cases is a guess wearing a uniform, which is why trees carry minimum-leaf-size settings.

The number and purity of leaves also describe the tree's complexity: a few pure leaves is a simple, explainable model; thousands of tiny leaves is memorization with a branching structure.

For a business, leaves are the readable unit of a tree model: you can open any leaf, see who lands there and what they get predicted, which is why tree-based reasoning survives in regulated decisions.

In ensembles, the idea multiplies: a forest is many trees, and a case's final prediction pools the leaves it lands in across all of them, but the individual leaf remains the atom of the reasoning.

How it works

Grow the tree by splitting, and stop each branch when further splits fail the rules: minimum leaf size, maximum depth, no gain.

Compute each leaf's prediction from its training cases: an average for numbers, a majority vote for classes.

Inspect leaf sizes and purities: tiny or mixed leaves are where the model is guessing.

Prune back branches whose leaves do not earn their complexity on validation data.

Report predictions with the leaf's population: a prediction from a large, pure leaf deserves more confidence than one from a small, mixed one.

Three practical examples

A churn tree whose largest leaf holds 40 per cent of customers at 3 per cent churn, and whose smallest flags a 60-case micro-segment at 70 per cent.
A pricing tree pruned so no leaf holds fewer than 200 historical sales, keeping every quote statistically respectable.
An audit that walks a disputed decision down its path and opens the leaf: who else lands here, and what happened to them.

Think of it this way

Think of a sorting office where mail cascades down chutes into final sacks. Each sack is a leaf: everything in it goes to the same destination, and the sack's contents tell you exactly what that destination's mail looks like.

The limit: a sack with three letters is a rumor, not a route. Predictions from thin leaves inherit the small-sample luck of whoever happened to land there, which is why the minimum-leaf setting is a guardrail, not a detail.

Worked example

opening the thin leaf

Suppose a lender's tree approves a borderline applicant, and the reviewer walks the path to the leaf. The figures are illustrative.

The approving leaf holds 23 historical cases, of which 16 repaid, a 70 per cent rate on a very small sample.

The reviewer's rule requires 100 cases per leaf for auto-approval, so the case routes to a human, who declines on other grounds.

Three months later the leaf's true rate settles near 55 per cent: the thin leaf had been a lucky draw, not a discovery.

Illustrative case study

A fictional retailer let its demand tree grow without a minimum leaf size, producing four thousand leaves, some holding single historical sales.

Forecasts for the thin leaves swung wildly week to week, orders whipsawed, and the supply chain spent a quarter unlearning the noise.

The invented lesson: every leaf is a claim about a population. Populations of eleven do not support claims; they support anecdotes.

Common misconceptions and mistakes

The leaf is the tree. The leaf is where one path's prediction lives; the model is the whole forest of paths and the discipline of the splits.
Bigger trees are smarter trees. Past a point, extra leaves slice noise, not knowledge, which is what pruning and minimum sizes exist to prevent.
Leaf predictions are precise. Each carries the sampling error of its population, and thin leaves are the least precise numbers in the model.

Business use

Set a minimum leaf size matched to the stakes, and enforce it in every tree-based model.

Report leaf populations with predictions in any decision that humans review.

Prune on validation, and keep the before-and-after evidence in the model file.

Check your understanding

Why does a leaf's population matter as much as its prediction?

Answer: Because the prediction is computed from the cases in that leaf, so a small population means a prediction built on a handful of anecdotes rather than a stable pattern.

Related terms

Learn next

AI for Business Owners explains how to read a tree model's output with a skeptical eye for thin evidence: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]

Sources

[S2] Google, Machine Learning Glossary. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Leabra Next: Leak node