One-class Ellipsoidal Kernel Machine for Outlier Detection

Differed from the bounding hypersphere assumption in Support Vector Machine (SVM), Ellipsoidal Kernel Machine (EKM) adopts the compacter bounding ellipsoid assumption, and finds the separating plane inside the ellipsoid. It reduces the VC dimension in essence. However, EKM only applies in binary classification and does not work in outlier detection where generally only one class of samples existed. Thus, this paper proposes a method for outlier detection One-class Ellipsoidal Machine and its kernel extension, which first finds a minimal ellipsoid enclosing all the input samples, and then finds the separating plane inside the ellipsoid by one-class SVM. Experiments on the artificial dataset and real datasets from UCI repository validate the effectiveness of the proposed method.