DataTerminology.comNEW
Also known as: Multi-Class Multi-Label, Multi-Output
A supervised learning task where each instance can be assigned multiple labels simultaneously.
DESCRIPTION
Multi-label classification extends traditional classification by allowing instances to belong to multiple classes at once. It is common in tagging problems like text categorization, image annotation, and music genre labeling. Models often use sigmoid outputs per label and thresholding for predictions. Key strategies include binary relevance, classifier chains, and adapted algorithms. Evaluation metrics include Hamming loss, micro/macro-averaged F1, and subset accuracy. Handling label imbalance, label correlations, and proper threshold tuning are central challenges.
ALTERNATIVE DEFINITIONS
Wikipedia
In machine learning, multi-label classification is the problem of assigning a set of labels to each instance.
Emphasizes the assignment of a set of labels per instance rather than a single class.
scikit-learn
In the multi-label classification problem, each sample can have more than one label associated with it.
Highlights multi-label as a supervised task where samples may belong to multiple classes.
RELATED TERMS
Binary Relevance Classifier Chains Hamming Loss F1 Score Label Imbalance One-vs-Rest Sigmoid Activation Thresholding Label Ranking
COMPETING TERMS
Multiclass Classification Single-Label Classification
RELATIONS
- Broader: Classification
KNOWLEDGE AREAS
TARGET AUDIENCE
Data Scientist AI Engineer MLOps Engineer Data Analyst AI Product Manager
EXTERNAL SOURCES
- Wikipedia - Multi-label classification
- scikit-learn - Strategies and APIs for multilabel tasks
- Google Developers - Practical guide to multilabel classification
- arXiv - Classifier chains seminal paper
TOP QUESTIONS
What makes Multi-Label Classification different from Multi-Class Classification?
Multi-class assigns exactly one label per instance, whereas multi-label allows multiple labels simultaneously. This changes model outputs, loss functions, evaluation metrics, and thresholding strategies.
Category: Misunderstanding
How do you evaluate Multi-Label Classification models?
Use metrics like Hamming loss, micro/macro-averaged precision, recall, F1, subset accuracy, and label ranking metrics. Micro vs. macro averaging addresses class imbalance differently.
Category: Testing
What are common strategies for Multi-Label Classification?
Binary relevance trains one classifier per label, classifier chains model label dependencies, and adapted algorithms modify models to output multiple labels directly.
Category: Methods
How should thresholds be chosen in Multi-Label Classification?
Calibrate per-label thresholds using validation data, optimizing metrics like F1 or precision/recall trade-offs. Global versus per-label thresholds can substantially affect performance.
Category: Techniques
What are common mistakes in Multi-Label Classification?
Using softmax instead of sigmoid, ignoring label imbalance, and evaluating with accuracy alone are frequent errors that degrade model utility.
Category: Mistakes
How do you handle label imbalance in Multi-Label Classification?
Apply class weighting, focal loss, resampling, or threshold tuning. Also consider macro-averaged metrics to reduce dominance by frequent labels.
Category: Requirements
When is Multi-Label Classification appropriate?
Use it when instances naturally belong to multiple categories, such as topic tagging, multi-genre music classification, or multi-ingredient recipe tagging.
Category: Application
What architectures work well for Multi-Label Classification?
Neural networks with sigmoid outputs per label, sequence models for text, and CNNs for images are common, often with multi-label loss like binary cross-entropy.
Category: Architecture
What is a hot topic in Multi-Label Classification?
Modeling label correlations explicitly, leveraging graph structures, and using prompt-based LLMs for zero-shot tagging are active research areas.
Category: Topic