DataTerminology.comNEW
A support vector method that learns a boundary around normal data to detect novel or outlier points.
DESCRIPTION
One-Class SVM models the support of a distribution, enclosing normal observations and labeling external points as anomalies. Using kernels like RBF, it captures non-linear boundaries in high-dimensional spaces. It is effective for novelty detection with clean training data containing only normal examples. Parameter nu controls the fraction of outliers and margin. Proper scaling and kernel selection are critical for robust performance.
ALTERNATIVE DEFINITIONS
scikit-learn
Estimate a support vector machine for novelty detection.
Positions the method specifically for novelty detection.
Journal of Machine Learning Research
We propose a method for estimating the support of a high-dimensional distribution.
Introduces the theoretical foundation for One-Class SVM.
RELATED TERMS
Data Anomaly Detection Novelty Detection Kernel Trick RBF Kernel Hyperparameter Tuning Decision Function Support Vector Machine
COMPETING TERMS
Isolation Forest Local Outlier Factor SVDD
KNOWLEDGE AREAS
TARGET AUDIENCE
Data Scientist AI Engineer MLOps Engineer
EXTERNAL SOURCES
- Wikipedia - Support-vector machine#One-class SVM
- PyOD - Python toolbox for outlier detection
- LIBSVM - SVM library with one-class support
TOP QUESTIONS
When should One-Class SVM be used?
Use it when only normal data is available and you need to flag novel points in deployment.
Category: Overview
What is a common mistake when training One-Class SVM?
Including contaminated outliers in training can distort the boundary and degrade detection.
Category: Mistakes
How does nu affect One-Class SVM?
Nu sets an upper bound on the fraction of anomalies and a lower bound on support vectors, tuning sensitivity.
Category: Requirements
How does scaling impact One-Class SVM performance?
Feature scaling strongly affects kernel distances; standardize inputs to prevent dominance by high-variance features.
Category: Best practices
What are alternatives to One-Class SVM?
Isolation Forest, Local Outlier Factor, and autoencoders are popular alternatives with different assumptions.
Category: Alternatives
What risks exist in deploying One-Class SVM?
Drift and changing data distributions degrade accuracy; monitor and retrain periodically.
Category: Risks
How to evaluate One-Class SVM without labeled anomalies?
Use synthetic anomalies, reconstruction error comparisons, or domain feedback metrics like alert precision.
Category: Testing
What kernel is commonly used in One-Class SVM?
RBF is typical due to flexibility and smooth decision boundaries.
Category: Opinion
Can One-Class SVM be used for fraud detection?
Yes, it can flag unusual transactions if trained on normal behavior and monitored for drift.
Category: Example