DataTerminology.comNEW
Also known as: OvO
A multiclass strategy that trains one binary classifier for each pair of classes and combines their decisions.
DESCRIPTION
One-vs-One (OvO) decomposes multiclass problems into pairwise binary classifiers. For K classes, it trains K(K−1)/2 classifiers, each distinguishing a class pair. Predictions aggregate votes or apply tournament-like schemes. OvO is efficient for algorithms like SVMs, often outperforming One-vs-Rest when classes are separable. It can be robust to class imbalance but increases training cost as classes grow.
ALTERNATIVE DEFINITIONS
scikit-learn
One-vs-one: constructs one classifier per pair of classes.
Defines OvO succinctly within scikit-learn’s multiclass strategies.
Wikipedia
In one-vs-one classification, one trains N(N−1)/2 binary classifiers, each trained on data from two classes.
Highlights the number of classifiers and pairwise training.
Stanford CS229
In one-vs-one, we train a separate classifier for each pair of classes and use voting to predict the final class.
Explains typical voting aggregation for OvO.
RELATED TERMS
Multiclass Classification Support Vector Machine Logistic Regression Error-Correcting Output Codes Ensemble Methods Class Imbalance Cross-Validation Hyperparameter Tuning
COMPETING TERMS
KNOWLEDGE AREAS
TARGET AUDIENCE
Data Scientist AI Engineer MLOps Engineer
EXTERNAL SOURCES
- Wikipedia - Multiclass classification
- scikit-learn - OvO and OvR documentation
- LIBSVM - Multiclass SVM strategies
- Google - Machine learning crash course
TOP QUESTIONS
When is One-vs-One preferable to One-vs-Rest?
OvO can be better when classes are well separated pairwise and base classifiers handle small datasets efficiently, as with SVMs.
Category: Opinion
How do OvO predictions aggregate?
Typically via majority voting among pairwise classifiers. Alternatives include tournament schemes or probability coupling.
Category: Overview
What are OvO’s computational trade-offs?
More classifiers increase training cost, but each model trains on fewer samples, often improving runtime with certain algorithms.
Category: Risk
How does class imbalance affect OvO?
OvO reduces imbalance impact per pairwise problem, but careful sampling or weighting may still be necessary.
Category: Mistakes
How to tune OvO systems?
Select strong base learners, tune hyperparameters per pair, and use cross-validation with appropriate aggregation strategies.
Category: Requirements
What is the biggest misunderstanding about OvO?
Some assume it always requires more compute. In practice, pairwise datasets can be smaller and faster for certain learners.
Category: Misunderstanding
What’s a starting point for implementing OvO?
Use scikit-learn’s OneVsOneClassifier wrapper with SVM or logistic regression, and evaluate with cross-validation.
Category: Learning
Which models pair well with OvO?
SVMs, linear models with strong regularization, and tree-based ensembles can perform well with OvO aggregation.
Category: Example
What are alternatives to One-vs-One?
One-vs-Rest, direct multinomial models, and error-correcting output codes are common alternatives.
Category: Alternatives