Classification

Classification is the task of predicting a category: the answer is one label from a fixed set of options.

Sorting the post is classification. Every envelope goes into exactly one pigeonhole — bills, personal, advertising — and the sorter's skill is reading the signals on the envelope: sender, stamp, handwriting. A classifier does the sorting from features, and its answer is always one of the pigeonholes it was trained on. It cannot invent a new one.

The flavours are named by counting:

binary       two options            spam / not spam
multi-class  one of many            which of 10 digits is this?
multi-label  several at once        a film tagged action AND comedy

Under the hood, most classifiers do not output a category directly. They output a score or probability per class — 0.93 spam, 0.07 not — and a threshold turns scores into a decision. That threshold is a business choice, not a model property: a bank screening fraud will accept more false alarms than an email filter should.

Its sibling task is regression, where the answer is a number on a scale rather than a pigeonhole. Predicting whether a loan defaults is classification; predicting how much will be recovered is regression. Evaluation has its own traps — start with the confusion-matrix and not raw accuracy.

Where to go next#

• Full lesson: Classification
• Related terms: regression, supervised-learning, logistic-regression, confusion-matrix