multilabel

Multilabel learning is a machine learning task in which each instance can be associated with multiple labels from a fixed set. This differs from multiclass classification, where exactly one label is assigned per instance. Multilabel problems occur in image and video tagging, document categorization, music annotation, and biology, where objects often bear several attributes simultaneously.

Data are represented as feature vectors with an accompanying label subset. A common representation uses a binary indicator matrix for labels. Methods fall into two broad groups: problem transformation and algorithm adaptation. Problem transformation converts the problem to multiple single-label tasks or a single multiclass task. Examples include binary relevance (one binary classifier per label), classifier chains (modeling label dependencies along a chain), and the label powerset approach (treating each distinct label combination as a class). Algorithm adaptation methods extend learners to predict several labels directly, such as multilabel-capable neural networks, decision trees, or k-nearest neighbors.

Evaluation uses both instance- and label-level metrics. Hamming loss, subset accuracy, and Jaccard similarity are common. Precision, recall, and F1 can be computed per label (macro) or overall (micro). Cross-validation should preserve label distributions to account for imbalance and label correlations.

Typical tools include scikit-learn, which provides multilabel capabilities and metrics, and dedicated libraries such as Mulan for Java. Datasets vary in label cardinality and density, and practical work often addresses label dependencies, imbalance, and scalable training for large label spaces.