Beyond yes-or-no classification, two common variations arise.
Multi-Class
Each example belongs to exactly one of several classes — classifying a support ticket as billing, technical or account.
Approaches:
- Models that handle multiple classes natively, like trees, forests and neural networks with softmax outputs.
- One-vs-rest: train one binary classifier per class.
- One-vs-one: train classifiers for each pair of classes.
Multi-Label
Each example can have several labels at once — an article tagged "health", "policy" and "research".
Approaches:
- Independent binary classifiers per label.
- Neural networks with sigmoid outputs per label.
- Methods that model label dependencies, such as classifier chains.
Metrics
- Multi-class: accuracy, per-class precision and recall, confusion matrices, and macro or weighted averages.
- Multi-label: per-label metrics, micro and macro averages, Hamming loss and subset accuracy.
Macro Versus Micro
Macro averaging treats every class equally, highlighting rare-class performance. Micro averaging weights by frequency.
Imbalance
Rare classes are common in both settings; check per-class results.