What does multi-class mean?

Each item gets exactly one label from a set of three or more possible categories.

How does it differ from multi-label?

Multi-label allows several labels per item; multi-class insists on exactly one.

What is a good accuracy in multi-class?

Judge it against the random rate for your class count and against the cost of the specific confusions.

What tool reveals multi-class errors?

The confusion matrix, which shows which classes are mistaken for which.

How many classes can a model handle?

From three to thousands, but the data floor, confusion surface and maintenance cost all grow with the count.

AI Glossary / AI and machine learning / Beginner

Multi-class

Definition

Multi-class describes a classification setting where each item belongs to exactly one of three or more categories: a ticket routed to one of ten teams, a product assigned to one of fifty categories.

What it means

The word marks the step past binary. Two classes, yes or no, is the simple case; multi-class keeps the single-answer rule but widens the menu, and most real routing problems live there.

Single membership is the defining property. A support ticket gets one team, a document gets one topic, a transaction gets one status. If an item can hold several labels at once, that is multi-label, a different setting.

Multi-class changes the evaluation. Accuracy still counts right answers, but errors now have structure: confusing billing with payments is different from confusing billing with legal, and the confusion matrix becomes the honest report.

It also changes the baseline: with ten balanced classes, guessing scores ten per cent, so a fifty per cent accuracy that would be poor in binary work can be five times the trivial rate here.

For a business, the practical points are the category list and the confusion structure: a sane, stable set of classes matters more than the algorithm, and the worst confusions are the ones that cost money.

Two classes is the special case with its own name, binary, and its own habits; multi-class is everything from three upward, where those habits stop working.

How it works

Define the category list carefully: mutually exclusive, collectively covering, and stable enough to train and evaluate against.

Check class balance, since multi-class problems often hide one dominant class and several rare ones.

Train a classifier that outputs one class per item, using a model or a combination of binary models under the hood.

Evaluate with the confusion matrix and per-class metrics, because the average hides which classes fail and how.

Review the worst confusions with domain experts, since some mix-ups are harmless and others are the ones that lose customers.

Keep an explicit policy for items that fit no class, since a forced choice teaches the model to invent memberships.

Three practical examples

A ticket router assigning each message to exactly one of twelve support queues.
A product catalog assigning each item to one of three hundred categories for browsing and search.
A language identifier labeling each customer message with one of twenty languages before routing it to a fluent agent.

Think of it this way

Think of a mailroom: every letter goes into exactly one pigeonhole. More holes means more chances to misfile, and some wrong holes, the invoice in the legal tray, matter far more than others.

Worked example

the fifty-category shelf

Suppose a marketplace must assign products to one of fifty categories. The figures are illustrative.

A model reaches 62 per cent accuracy, which sounds weak until set against the two per cent rate of guessing and the 40 per cent rate of the old keyword rules.

The confusion matrix shows most errors land between sibling categories, charger versus cable, which shoppers forgive, while the rare cross-department errors are the ones that break search.

The fix is not a new model but a taxonomy review: merging sibling categories nobody can reliably separate lifts accuracy and simplifies every downstream metric.

Illustrative case study

A fictional bank added classes to its complaint classifier every quarter as new products launched, reaching forty classes without ever re-baselining.

Accuracy slid from 85 to 60 per cent, which leadership read as model decay, but the task itself had quietly become harder and the old baselines no longer applied.

The invented lesson: in multi-class work, the category list is part of the model. Change the menu and every historical number changes meaning.

Common misconceptions and mistakes

Multi-class means an item can have several classes. That is multi-label; multi-class means exactly one class from three or more.
Accuracy reads the same as binary. The trivial rate drops as classes multiply, so the same accuracy can be bad in binary and strong in multi-class.
All errors are equal. Confusions have structure and cost; the matrix, not the average, is the report that shows them.

Business use

Keep the category list small, exclusive and stable, and treat changes to it as model changes.

Report accuracy against the random-guess rate for the actual number of classes.

Read the confusion matrix in reviews and price the expensive confusions explicitly.

Merge categories that neither the model nor your own staff can reliably separate.

When comparing two models, compare them on the same category list, since a taxonomy change makes the numbers incomparable.

Check your understanding

Why does the number of classes change how accuracy should be read?

Answer: Because the random-guess rate falls as classes multiply, so a given accuracy means something different in binary, ten-class and hundred-class settings, and only comparison with the trivial rate for that class count gives the number meaning.

Related terms

Learn next

Prompt Like a Pro helps you read classifier reports and ask for the matrix behind the number: https://aiforbusinessandlife.com/academy/prompt-like-a-pro-clear-framework. [C3]

Sources

[S2] Google, Machine Learning Glossary; [S24] scikit-learn documentation, Glossary of Common Terms and API Elements. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Multi-attribute global inference of quality Next: Multi-class classification