One Class SVM Anomaly Detection

One-class SVM is an unsupervised anomaly detection method that learns a boundary enclosing most of the training data in a kernel feature space and flags observations outside that boundary as anomalies.

원클래스 SVM은 커널 특성공간에서 학습 데이터의 대부분을 둘러싸는 경계를 학습하고 그 경계 밖의 관측치를 이상치로 표시하는 비지도 이상탐지 방법입니다.

How do I detect unusual observations when I have no labelled examples of anomalies?

이상치에 대한 라벨이 없을 때 비정상적인 관측치는 어떻게 탐지하나요?

One-class SVM learns the region occupied by typical observations and labels points that fall outside it as anomalies, with the nu parameter setting an upper bound on the fraction treated as outliers. It fits small to medium datasets of scaled numeric features where the training data are mostly normal and anomalies are not labelled.

When to use One Class SVM Anomaly Detection

• Mostly normal numeric data with no labelled anomalies
• Fraud, fault, or quality screening on small to medium datasets
• Nonlinear boundaries between normal and unusual cases

When to use something else

• Large or high-dimensional datasets: use isolation forest
• Labelled examples of anomalies are available: use a supervised classifier

Assumptions and requirements

• Numeric features scaled to comparable ranges
• Training data dominated by normal observations
• Kernel parameter (gamma) and nu tuned with care
• Computation grows quickly with sample size

What the analysis reports

• Anomaly label for each observation
• Decision function scores
• Proportion of observations flagged
• Plot of scores or a low-dimensional projection highlighting anomalies

How to report it

The one-class SVM (RBF kernel, ν = [value]) flagged [value]% of observations as anomalies.

Related methods

• Isolation Forest
• Local Outlier Factor
• Mahalanobis Distance Outlier Detection

Run One Class SVM Anomaly Detection on Quorant