Mdl = ocsvm (X) fits a one-class support vector machine to the $N$-by-$P$ matrix X, whose rows are observations and columns are variables, and returns a OneClassSVM object Mdl.

[Mdl, tf, scores] = ocsvm (X) also returns the $N$-by-1 logical vector tf flagging the anomalous observations and the $N$-by-1 vector scores of anomaly scores. A higher score indicates an observation that lies further outside the boundary enclosing the data, and is therefore more likely to be an anomaly.

The observations are mapped to a randomized feature space that approximates a Gaussian kernel of scale KernelScale using NumExpansionDimensions features, and a linear one-class boundary is fitted there with ridge regularization of strength Lambda.

Additional parameters can be specified by Name-Value pair arguments.

The feature expansion uses random projections, so the scores depend on the state of the random number generator and are not reproducible across runs unless the generator is seeded. The 'auto' selections and the fitted model differ from MATLAB’s implementation, which uses a different feature expansion and solver.

For a deterministic, classic one-class support vector machine (a nu-SVM with an exact kernel, computed through libsvm), use fitcsvm with a single class in the response or with the 'Nu' name-value argument; that path returns a ClassificationSVM object whose predict method labels observations, rather than the anomaly-scoring interface provided here.

Use the isanomaly method of Mdl to detect anomalies in new data.

See also: OneClassSVM, isanomaly, iforest, lof, fitcsvm

Source Code: ocsvm