Definition
Classification assigns one or more discrete categories to an input. A classifier often produces a score or estimated probability for each class, then a decision rule converts those values into labels.
The class structure is part of the problem definition. In single-label classification, one class is chosen from mutually exclusive options. Multi-label classification allows several labels at once. Hierarchical classification organizes labels into levels, such as animal, bird, and hawk. Binary classification is the two-class case, even when the implementation exposes only the score for the positive class.
What determines a useful classifier
Accuracy alone can hide the error that matters. The class prevalence, decision threshold, false-positive cost, false-negative cost, calibration, and abstention policy determine how a score becomes an operational decision. A classifier can rank examples well and still perform poorly after a threshold is chosen for the wrong cost tradeoff.
Classes are also human choices. Ambiguous label instructions, changing policies, annotator disagreement, and missing categories can cap performance before model selection begins.
Distinguish it from nearby terms
- Regression predicts a continuous quantity rather than a category.
- Clustering groups examples without starting from predefined class labels. A person still has to interpret what the groups mean.
- Ranking orders candidates. A later rule may classify the highest-ranked items, but ordering and labeling are different tasks.
- Detection finds and localizes instances, such as objects in an image. It often includes classification as one step.
Check your understanding
A fraud model returns a score of 0.62. The classification is not fully specified until the team defines the threshold, what happens near it, and the cost of blocking a legitimate transaction versus allowing fraud.