What is multilabel classification?

What is multilabel classification?

• A. A classification problem where multiple labels can be true at the same time for one instance.
• B. A single label per instance.
• C. Forecasting probabilities for classes.
• D. A clustering approach.

Answer: (A)

Multilabel classification means each instance can belong to multiple labels at the same time. Instead of picking a single category for an item, the model outputs a set of applicable labels, often represented as a binary vector where each position indicates whether a label is present. This situation appears in tasks like image tagging (an image may be labeled “beach,” “people,” and “sunset” all together) or news topic tagging (an article can cover several topics simultaneously). This setup matters because evaluation and modeling acknowledge that more than one label can be true, so metrics and methods differ from single-label classification. Common evaluation measures include per-label accuracy and loss (like Hamming loss), as well as metrics that summarize performance across all labels such as micro- and macro-averaged precision, recall, and F1, and sometimes subset accuracy which requires all labels for an instance to be exactly correct. Other options don’t fit because forecasting probabilities for classes is about the model’s outputs rather than the structural possibility of multiple labels, and clustering is an unsupervised grouping task, not a supervised multi-label assignment. A single label per instance describes a different, simpler setup.

Multilabel classification means each instance can belong to multiple labels at the same time. Instead of picking a single category for an item, the model outputs a set of applicable labels, often represented as a binary vector where each position indicates whether a label is present. This situation appears in tasks like image tagging (an image may be labeled “beach,” “people,” and “sunset” all together) or news topic tagging (an article can cover several topics simultaneously).

This setup matters because evaluation and modeling acknowledge that more than one label can be true, so metrics and methods differ from single-label classification. Common evaluation measures include per-label accuracy and loss (like Hamming loss), as well as metrics that summarize performance across all labels such as micro- and macro-averaged precision, recall, and F1, and sometimes subset accuracy which requires all labels for an instance to be exactly correct.

Other options don’t fit because forecasting probabilities for classes is about the model’s outputs rather than the structural possibility of multiple labels, and clustering is an unsupervised grouping task, not a supervised multi-label assignment. A single label per instance describes a different, simpler setup.