Usually one head that scores all classes and picks the highest, or several binary models whose verdicts combine.

What data does it need?

Examples labeled with their single true class, with enough per class for even the rare ones to be learned.

Overall accuracy for the headline, the confusion matrix and per-class recall for the truth.

When should low confidence go to a human?

When the cost of a forced wrong guess exceeds the cost of review, which for rare or regulated classes is nearly always.

AI Glossary / AI and machine learning / Beginner

Multi-class classification

Q: What is multi-class classification?

The task of assigning each item to exactly one category from three or more options.

Definition

Multi-class classification is the task of assigning each item to exactly one category from three or more: routing a ticket to one of ten teams, tagging a document with one of twenty topics.

What it means

This is the workhorse task behind most routing and sorting automation. Wherever items arrive and must be filed into one of several bins, multi-class classification is the formal name of the job.

Models handle it two ways: natively, with one head that scores every class and picks the highest, or by reduction, training several binary models and combining their verdicts. Modern practice favors the native head.

Training data needs the same structure: every example labeled with its single true class, and enough examples per class that even the rarest bins are learnable.

Evaluation lives in the confusion matrix: overall accuracy for the headline, per-class recall for the weak spots, and the specific confusion pairs for the decisions that cost money.

For a business, the deliverable to demand is not a score but the routing table: where do errors go, which bins starve, and what does the worst misrouting cost compared with a human doing the filing.

The setting appears under other names in vendor material: single-label multi-class, multinomial classification, and plain classification when the context implies several bins.

How it works

Set the category list with the business: exclusive, complete and stable, with an 'other' bin policy decided in advance.

Collect labeled examples per class, targeting a floor for the rarest class, not just a total count.

Train a classifier with a single multi-class head or a set of binary reductions, validating on held-out data.

Read the confusion matrix per class: recall per bin, the dominant confusion pairs, and their business cost.

Set fallbacks for low-confidence items, routing them to human review rather than forcing a bad guess into a bin.

Monitor per-class volume in production, because a class that stops appearing usually signals a process change, not a quiet model.

Three practical examples

An email triage assigning each message to sales, support, billing or careers, with low-confidence messages going to a human queue.
A document system filing contracts, invoices and correspondence automatically, with per-class recall watched monthly.
A moderation tool assigning content to one of several policy categories, where some confusions carry legal weight.

Think of it this way

Think of sorting mail by hand into department trays: one letter, one tray, every day. Automating that is multi-class classification, and the quality question is which trays get crossed, not just how often.

Worked example

the rare bin that starved

Suppose a telco routes complaints into eight classes, and the legal-threat class is two per cent of volume. The figures are illustrative.

The model scores 84 per cent overall, but per-class recall shows legal threats caught only 30 per cent of the time, invisible in the headline.

Targeted collection adds four hundred labeled legal examples, class weights are applied, and legal recall doubles at a cost of two points of overall accuracy.

The trade is the right one: the rare bin carried the highest cost per miss, and the confusion matrix, not the average, was what made the case for spending on it.

Illustrative case study

A fictional logistics firm deployed a nine-class parcel-issue router and celebrated 88 per cent accuracy while the dangerous-goods class, under one per cent of volume, was caught barely half the time.

A misrouted dangerous-goods report reached a general queue and sat for two days, producing the incident that rewrote the evaluation process.

The invented lesson: in multi-class classification, publish per-class recall next to every headline score, and let the costliest class set the acceptance bar.

Common misconceptions and mistakes

It is just binary classification repeated. It is one task with its own training heads, baselines and confusion structure.
Overall accuracy is the quality measure. It is the average over classes with wildly different costs; per-class recall is the serious report.
More classes are free. Every added class raises the data floor, the confusion surface and the maintenance bill.

Business use

Fix the category list before modeling, and budget labeled data for the rarest class, not the average one.

Acceptance-test on per-class recall for the costliest classes, not on overall accuracy.

Add a low-confidence route to human review instead of forcing every item into a bin.

Re-baseline all historical metrics whenever the category list changes.

Check your understanding

Why is per-class recall the serious report in multi-class classification?

Answer: Because overall accuracy averages across classes with different volumes and costs, so a model can score well while failing the rare, expensive class, and only per-class recall shows each bin's real catch rate.

Related terms

Learn next

Prompt Like a Pro shows how to request per-class evidence instead of headline scores: https://aiforbusinessandlife.com/academy/prompt-like-a-pro-clear-framework. [C3]

Sources

[S2] Google, Machine Learning Glossary; [S19] scikit-learn documentation, Metrics and scoring. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Multi-class Next: Multi-class logistic regression