What kinds of questions can a tree ask?

For numbers, threshold questions such as above or below fifty; for categories, membership questions such as channel is email or not.

Can a tree predict numbers?

Yes. Regression trees store an average in each leaf instead of a majority category.

Why did my tree change after I added a little data?

Trees are unstable learners: small data changes can flip an early split, and everything below that split changes with it.

AI Glossary / AI and machine learning / Beginner

Decision tree

Definition

A decision tree is a model that predicts an answer through a sequence of yes-or-no style questions about the input, arranged like the branches of a tree, ending at a leaf that holds the prediction.

What it means

Each internal point of the tree asks one question about one field, such as whether the customer's order value exceeds fifty pounds, and the answer sends you down one of two branches.

Following the branches to the end lands you at a leaf. The leaf holds the prediction: a category such as will renew, or a number such as expected spend, learned from the training examples that fell down the same path.

The appeal is transparency. You can draw the tree, read off the exact rule that produced any answer, and discuss it with people who have never heard of machine learning.

The weakness is fragility. A deep tree can memorize its training data, quirks included, and a small change in the data can grow a different tree, so single trees often predict poorly on fresh cases.

For a business, trees are the natural first model when the explanation matters as much as the answer, and they are the building block inside the far stronger forest methods.

How it works

Start with all training examples at the root and search every field for the yes-or-no split that best separates examples with different outcomes.

Split the examples into two groups along that best question, then repeat the same search inside each group, growing branches downward.

Stop growing when a group is too small, too uniform, or too deep, and declare it a leaf holding the majority answer or average of its examples.

Optionally prune the grown tree back, removing branches whose detail only fits noise, which improves performance on new data.

To predict, route the new case down the branches by answering each question, and return the answer stored in the leaf it reaches.

Three practical examples

A telecom tree that asks about contract length, monthly spend and support calls to predict which customers will churn this quarter.
A clinic triage tree that sequences questions about symptoms to route patients to urgent, routine or self-care pathways.
A pricing tree that predicts the discount a deal needs, based on deal size, sector and time in the pipeline.

Think of it this way

Think of the party game twenty questions. Each question halves the possibilities, and a short run of good questions pins down the answer. A decision tree is twenty questions where the questions were chosen in advance by studying past games.

The limit: if the past games were unusual, the pre-chosen questions are the wrong ones, and the game falls apart on new players.

Worked example

reading a churn tree

Suppose a telecom grows a tree on 10,000 past customers, with churn as the outcome. The figures are illustrative.

The root splits on contract type: monthly customers churn at 22 per cent, annual at 6 per cent. Among monthly customers, the next split is on support calls: three or more calls in a quarter lifts churn to 41 per cent.

A path down the tree is a rule staff can read aloud: monthly contract plus heavy support contact means high risk, and the leaf holds the 41 per cent figure.

The team uses those rules to brief retention staff, while checking the tree's accuracy on customers it never saw during training.

Illustrative case study

A fictional insurer grew a very deep tree for claims triage that scored beautifully on past claims and was celebrated internally.

On the next quarter's claims it performed no better than a coin toss: the deep branches had memorized accidents of the training data, such as a coincidence between postcode and payout size.

Pruning the tree to a third of its depth cut training accuracy slightly but restored real-world performance. The invented lesson: a tree that knows the past too well usually knows the future too little.

Common misconceptions and mistakes

A decision tree is only for experts. Reading a tree needs no mathematics at all; that is its main virtue.
A bigger tree is a better tree. Beyond a point, extra depth fits noise, and smaller trees predict new data better.
The tree gives reasons. It gives rules that correlate with the outcome; whether a field causes the outcome is a separate question the tree cannot answer.

Business use

Use a single tree when the explanation will be read by customers, regulators or front-line staff; its transparency is a genuine feature.

Treat a single tree's accuracy with suspicion and always measure it on held-out data; if you need accuracy more than readability, move to a forest.

Review which fields the tree splits on near the root: those are the drivers your stakeholders will ask about, and surprises there often reveal data problems.

Check your understanding

Why is a shallow decision tree often preferred over a deep one for real decisions?

Answer: Deep branches tend to fit accidents of the training data, so a shallower tree usually predicts new cases better and is far easier to explain.

Related terms

Learn next

AI for Business Owners explains how to read a model's rules and spot when an impressive score hides a fragile model: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]

Sources

[S1] Russell and Norvig, AIMA glossary; [S2] Google, Machine Learning Glossary; [S12] Aalto University, Dictionary of Machine Learning; [S14] KNIME, Data Science Glossary. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Decision theory Next: Decision tree pruning