Ensemble One-Class SVM
Ensemble of One-Class Support Vector Machines · Also known as: Ensemble OC-SVM, multiple one-class SVM, OC-SVM ensemble, one-class SVM committee
Ensemble One-Class SVM combines multiple one-class support vector machine models — each trained on a different random subset of the data or features — and aggregates their anomaly scores. By pooling several OC-SVM boundary estimates, the ensemble reduces the sensitivity to kernel choice and data sampling that afflicts a single one-class SVM, producing a more stable and accurate novelty or outlier detector.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use Ensemble One-Class SVM when only normal-class data are available for training (novelty detection) or when the dataset is strongly imbalanced with very few labelled anomalies (outlier detection). It outperforms a single OC-SVM when the data are high-dimensional, when the optimal kernel or nu parameter is uncertain, or when robustness to training-set variation is required. Do not use it when fully labelled two-class data are available — standard supervised classifiers will perform better. Avoid it on very small datasets where bootstrapping produces nearly identical base models and the ensemble overhead is not justified. For interpretability requirements, prefer simpler anomaly scores such as isolation forest with SHAP.
Strengths & limitations
- Reduces the high variance of a single one-class SVM caused by kernel and hyperparameter sensitivity.
- Aggregated scores provide smoother, better-calibrated anomaly rankings than a single decision boundary.
- Works entirely from normal-class data — no labelled anomalies required for training.
- Diversity strategies (random subspace, bootstrap) allow straightforward parallelisation of base model training.
- Generalises better to unseen anomaly types compared to a single OC-SVM fitted to one subregion of feature space.
- Computationally expensive: fitting many SVM models scales poorly to very large datasets without approximate or kernel approximation methods.
- Selecting diversity strategy and the number of base models adds hyperparameter complexity beyond a single OC-SVM.
- Interpretability is very low — neither the ensemble score nor any individual OC-SVM decision boundary has an intuitive explanation.
- Performance gains over a single well-tuned OC-SVM or over isolation forest are not guaranteed on all datasets.
Frequently asked
How many base OC-SVM models should the ensemble contain?
Typical ensembles use between 10 and 50 base models. Gains diminish quickly beyond 30 to 50 models, and the computational cost scales linearly with the number of base SVMs fitted. Start with 20 and monitor held-out AUROC as you add more.
How does this compare to isolation forest for anomaly detection?
Isolation forest is generally faster, scales better to large datasets, and is easier to tune. Ensemble OC-SVM may outperform isolation forest when the normal region has a compact kernel-representable shape, but for most general-purpose anomaly detection tasks isolation forest or LOF are preferred starting points.
What kernel should I use for the base OC-SVMs?
The RBF (Gaussian) kernel is the standard default. Vary the gamma parameter across base models to introduce diversity; alternatively, use a random feature approximation of the RBF kernel for scalability.
Can I use this method when I have some labelled anomalies?
If you have even a small number of labelled anomalies, semi-supervised methods or standard ensemble classifiers with appropriate class weighting will typically outperform a purely one-class approach. Reserve one-class methods for truly unlabelled or single-class settings.
How do I set the contamination rate?
Set it based on domain knowledge (e.g., historical fault rate) or estimate it from the training data distribution. Avoid using test-set information to select the threshold, as this produces overly optimistic detection rates.
Sources
- Scholkopf, B., Platt, J. C., Shawe-Taylor, J., Smola, A. J., & Williamson, R. C. (2001). Estimating the support of a high-dimensional distribution. Neural Computation, 13(7), 1443–1471. DOI: 10.1162/089976601750264965 ↗
- Tax, D. M. J., & Duin, R. P. W. (2001). Combining one-class classifiers. In Multiple Classifier Systems (MCS 2001), Lecture Notes in Computer Science, vol 2096. Springer, Berlin, Heidelberg. DOI: 10.1007/3-540-48219-9_30 ↗
How to cite this page
ScholarGate. (2026, June 3). Ensemble of One-Class Support Vector Machines. ScholarGate. https://scholargate.app/en/machine-learning/ensemble-one-class-svm
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Autoencoder Anomaly DetectionMachine learning↔ compare
- Isolation ForestMachine learning↔ compare
- One-class SVMMachine learning↔ compare
- Voting EnsembleMachine learning↔ compare