What is a node in a decision tree?

One point in the tree: internal nodes test a condition, leaf nodes carry the final prediction.

What does a leaf node store?

The prediction for items reaching it, usually the majority class or average value of its training examples.

Deep nodes memorize training noise; pruning keeps only the splits supported by enough examples to generalize.

How do nodes explain a prediction?

As the path: the exact sequence of questions and answers that carried the item from the root to its leaf.

AI Glossary / AI and machine learning / Intermediate

Node (decision tree)

Definition

A node in a decision tree is one decision point or conclusion: internal nodes test a condition on a feature, and leaf nodes carry the final prediction.

What it means

A tree is a flowchart of questions, and nodes are its boxes. Each internal node asks one question, is income above this value, has the customer called twice, and each answer sends the item down one branch toward a leaf.

The leaf is where the model speaks: the prediction stored there is usually the majority class or the average value of the training examples that arrived at that leaf, so each leaf is a tiny local model.

Reading a tree means reading its nodes: the question asked, the split chosen, and how many examples pass through, which is why trees are the most inspectable models in common use.

Nodes are also where trees overfit: split deep enough and every leaf memorizes a handful of examples, so pruning, limiting depth or requiring minimum leaf sizes is the discipline that keeps nodes honest.

For a business, the node is the unit of explanation: any single prediction can be narrated as the exact sequence of questions that produced it, which regulators and customers can both follow.

How it works

Start at the root node, which holds all training data and the first chosen split.

At each internal node, test the item on the node's condition and send it down the matching branch.

Repeat until a leaf node is reached, and read the prediction stored there.

Grow nodes during training by choosing splits that best separate the outcomes, until stopping rules halt the growth.

Prune or constrain nodes, depth limits, minimum leaf sizes, so leaves represent enough examples to be believed.

Log the example counts at every node during training review, since thin nodes are where confident nonsense grows.

Three practical examples

A credit tree whose root asks about missed payments, splitting the whole book into two very different risk populations.
A churn tree's leaf holding eighty customers, of whom sixty per cent left, which becomes the retention team's target list.
A pruned tree with forty nodes replacing an unpruned one with four thousand, at a cost of one accuracy point and a gain of total legibility.

Think of it this way

Think of twenty questions: each question is a node, each answer narrows the room, and the final guess is the leaf. A good game, like a good tree, asks the questions that halve the room best.

Worked example

the leaf with four examples

Suppose an unpruned churn tree grows until one leaf holds four customers, all of whom churned. The figures are illustrative.

The leaf predicts churn with total confidence, and the path to it reads persuasively: region, tenure, plan, calls. But four examples are a coincidence, not a rule.

Requiring leaves of at least fifty examples collapses that branch, and the pruned tree loses a point of training accuracy while gaining three on held-out data.

The node count tells the story: the deep tree had a thousand nodes memorizing noise, the shallow one has ninety encoding structure.

Illustrative case study

A fictional lender published its decision tree to prove transparency, but the tree was eleven levels deep with two thousand nodes, technically readable and practically unreadable.

Auditors found contradictory paths and leaves resting on single examples, and the transparency exercise became an embarrassment rather than a defense.

The invented lesson: a tree is explainable in proportion to its size. Prune until every node earns its place, or the flowchart is camouflage, not explanation.

Common misconceptions and mistakes

Every node is a conclusion. Only leaves conclude; internal nodes only ask questions and route.
Deeper trees know more. Deeper trees memorize more; knowledge is the splits that survive on new data, not the total count.
Nodes are chosen by rules of thumb. They are chosen by measured separation power on training data, which is why they must be re-earned on held-out data.

Business use

Read important predictions as node paths, and publish the path with the decision where customers or regulators are affected.

Constrain tree growth with minimum leaf sizes matched to how many examples a conclusion needs to be believed.

Prune to legibility: a tree your reviewers cannot hold in their heads is not the transparent model it claims to be.

Audit leaves for thin evidence, since confident leaves on few examples are where tree models lie.

Check your understanding

Why is a decision tree explainable only in proportion to its size?

Answer: Because explanation means reading the nodes: a small tree's questions and leaves can be inspected and narrated end to end, while a deep tree's thousands of nodes are technically readable but practically opaque, with confident leaves resting on thin evidence.

Related terms

Leaf
Decision tree
Pruning
root node
split
classification tree

Learn next

AI for Business Owners helps you read the models vendors call transparent: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]

Sources

[S12] Aalto University, Dictionary of Machine Learning; [S2] Google, Machine Learning Glossary. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Node Classification Next: Node degree