The SVDD Procedure

The SVDD procedure implements the support vector data description (SVDD) algorithm (Tax and Duin 2004). SVDD is a one-class classification technique that is useful in applications where data that belong to one class are abundant but data about any other class are scarce or missing. Fraud detection, equipment health monitoring, and process control are some examples of application areas where the majority of the data belong to one class. You can use SVDD to model such one-class data and subsequently use the model to perform outlier detection.

In its simplest form, an SVDD model is obtained by building a minimum-radius hypersphere around the one-class training data. The hypersphere provides a compact spherical description of the training data. This training data description can be used to determine whether a new observation is similar to the training data observations. The distance from any new observation to the hypersphere center is computed and compared with the hypersphere radius. If the distance is more than the radius, the observation is designated as an outlier. Using kernel functions in SVDD formulation provides a more flexible description of training data. Such description is nonspherical and conforms to the geometry of the data.

PROC SVDD implements only the flexible data description.

SVDD is useful for obtaining a geometric description of data and in most applications also for detecting outliers. SVDD is used in domains where the majority of the data belong to one class. Applications of SVDD include the following: