New one-class classifiers based on the origin separation approach

• Clear description of the origin separation approach
• Proof of the connection between ν-SVM and νoc-SVM
• New batch one-class classifiers with the origin separation approach
• New online one-class classifiers with the origin separation approach

The model of the one-class support vector machine (νoc-SVM) is based on the “origin separation approach,” i.e., to add a sample at the origin to the training data for the second class and apply a maximum margin separation as known from the classical SVM (C-SVM). This has been proven only for hard margin separation but a clearly defined relation between the νoc-SVM and the C-SVM is not yet existing. In this work, the origin separation approach is analyzed in more detail. The approach reveals to be a more general concept to relate binary and unary (one-class) classifiers. We prove how its application to the ν-SVM, a variant of the C-SVM, directly results in the νoc-SVM. Furthermore, we apply this concept to the C-SVM and other related methods (balanced relative margin machine, regularized Fisher’s discriminant analysis, online passive-aggressive algorithms) to derive entirely new classifiers. This includes variants that can be updated online which allows the application on large datasets or on systems with very limited resources.

Figure optionsDownload high-quality image (255 K)Download as PowerPoint slide