How is classification different from regression?

Classification picks a category; regression predicts a number.

How much labeled data do I need?

Enough to cover every class well, including rare and borderline cases; hundreds per class is a starting point, not a law.

What is a confusion matrix?

A table of predicted versus actual classes that shows exactly which categories get mixed up.

What is a baseline in classification?

A trivial strategy, such as always predicting the largest class, that any real model must beat to justify itself.

AI Glossary / AI and machine learning / Beginner

Classification

Definition

Classification is the task of assigning items to predefined categories, such as sorting emails into spam or not spam, using patterns learned from labeled examples.

What it means

Classification is the workhorse of applied AI. Routing tickets, screening transactions, tagging photos, qualifying leads: wherever items must be sorted into known buckets, a classifier does the sorting.

It is supervised learning, which means it learns from examples people have already labeled. The quality of those labels caps the quality of the model, quietly and completely.

It differs from its neighbors. Regression predicts a number, such as next month's revenue. Clustering finds groups nobody named in advance. Classification assigns names people chose.

Success is measured per decision, not per average. Accuracy counts right answers, but the confusion matrix shows which classes get mixed up, and the costs of those mix-ups are rarely equal.

For a business, a classification project is as much design as modeling: pick classes that map to actions, label examples consistently, set thresholds for the costs you face, and plan for the world to drift after launch.

Almost every AI feature a business buys has a classifier inside it, often several, deciding what an input is before anything clever is done with it.

The threshold question is always present: a classifier that is 95 percent sure can act alone, while one that is 55 percent sure should probably ask a person.

How it works

Define the classes and their boundary rules, then label a training set that covers the common cases and the awkward ones.

Train a model to score each class for each item, and hold out labeled examples the model never sees, so testing is honest.

Evaluate with per-class precision and recall and the confusion matrix, since the interesting failures hide inside the averages.

Set decision thresholds and review queues around the costs: cheap mistakes automated, expensive ones routed to people.

After launch, monitor accuracy and incoming-data drift, and retrain on fresh labels as products, language and behavior change.

Drift monitoring compares today's inputs and outcomes with the training era. When the mix of classes or the language shifts, performance usually follows, and fresh labels restart the loop.

Three practical examples

Email and ticket routing to the right team, with low-confidence items held for a person.
Quality inspection: photos of products classed as pass or defect on a production line.
Document processing: invoices, contracts and CVs sorted by type before the details are extracted.

Think of it this way

Think of training a new mail sorter by handing them a stack of pre-sorted letters. They study the stack, then sort fresh mail by the patterns they absorbed.

The limit: the sorter only knows the piles you showed them. A letter for a department nobody mentioned has nowhere sensible to go.

Worked example

auto-tagging a support inbox with a safety net

Suppose a firm gets 2,000 support emails a week, sorted by hand into six queues. Staff label 3,000 past emails, and a classifier is trained on them.

On held-out emails it tags 90 percent correctly, but review shows it is weak on the two smallest queues, exactly where a miss costs the most. Figures are illustrative.

The team routes low-confidence emails to humans, collects their corrections as new labels, and retrains monthly.

After three months, coverage and accuracy both rise, because the model learns from the very cases it found hardest.

Illustrative case study

A fictional manufacturer photographs every unit on the line and trains a classifier for pass, cosmetic defect and critical defect.

Early versions confuse cosmetic with critical more often than expected. The fix is not a bigger model but better labels: inspectors agree written boundary rules with photos, and relabel the disputed cases.

The invented lesson: in classification, arguments about labels are not a distraction from the work. They are the work.

The manufacturer also prices the errors before go-live: a missed critical defect costs a recall, a false alarm costs seconds, which is why the threshold favors catching defects.

Common misconceptions and mistakes

Classification and clustering are the same. Classification assigns known classes; clustering finds unnamed groups, a different job.
Perfect training accuracy means the model is ready. It usually means the model memorized; held-out data is the honest test.
A deployed classifier stays accurate. Language, products and behavior drift, so monitoring and retraining are part of the cost.

Business use

Choose classes that map to real actions, fund careful labeling including boundary cases, and insist on held-out, per-class evidence before rollout.

Budget for monitoring and periodic retraining from day one, and keep a human route for low-confidence and high-cost decisions.

Before buying classification as a feature, run a bake-off on your own labeled samples; a vendor's accuracy on their data says little about your queues.

Check your understanding

What is the difference between classification and clustering?

Answer: Classification assigns items to classes people defined and labeled in advance; clustering finds groups in unlabeled data without predefined names.

Related terms

Learn next

AI for Business Owners walks through framing, judging and rolling out prediction projects in a small business: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]

Sources

[S1] Russell and Norvig, AIMA glossary; [S2] Google, Machine Learning Glossary; [S3] NIST AI Resource Center glossary; [S12] Aalto University, Dictionary of Machine Learning; [S14] KNIME, Data Science Glossary; [S21] Data Science Glossary (datascienceglossary.org). Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Classical planning Next: Classification model