A photograph can show a bicycle, a person and a road at the same time. If the task is to identify all relevant contents, forcing the model to choose just one of those labels changes the task before training even begins.
Scikit-learn’s classification guide distinguishes multiclass classification, where an example receives one class from several possibilities, from multilabel classification, where several binary label decisions can apply to the same example.
Write the labelling rule in ordinary language
Consider a fictional support inbox. One message asks about a delayed delivery and a refund. A multiclass scheme might require a single routing category. A multilabel scheme could record both topics. Neither is automatically superior; they serve different operational questions.
Before choosing a model, decide whether labels are mutually exclusive, whether no label is allowed, and how ambiguous cases are handled. Include examples that challenge the rules. If annotators cannot agree what the task permits, changing the classifier will not resolve the underlying definition.
| Task | Allowed result |
|---|---|
| Multiclass routing | One destination from the available categories |
| Multilabel topics | Delivery and refund can both apply |
Represent absence carefully
A zero in a label column may mean the label is absent, but an unreviewed label is a different situation. Do not silently convert missing annotation into a confident negative example. Record the annotation process so the training data means what the model is asked to learn.
Check that the selected estimator and target representation support the task. A model’s ability to output several numbers does not by itself establish that the pipeline implements a suitable multilabel problem.
Evaluation also needs the same task definition. If two labels are correct and the model predicts one, decide how that partial result will be measured and what it means for the downstream workflow. Keep a few reviewed examples beside aggregate scores.
The first design decision is therefore conceptual: can one example legitimately belong to several classes? Answering that with concrete cases makes the later choices about labels, predictions and error handling much easier to defend.