D-optimality for Minimum Volume Ellipsoid with Outliers
Alexander Dolia, Scott Page, Neil White and Chris Harris
In: The Seventh International Conference on Signal/Image Processing and Pattern Recognition, UkrObraz-2004, 11-15 Oct 2004, Kiev, Ukraine.
A family of one-class classification methods is extended by the determinant maximization novelty detection (DMND) model based on the D-optimum experimental design approach for ellipsoif estimation. Similar to the one-class classification methods based on support vector machines or the so-called support vector data description (SVDD) approach, DMND is a method that fits a geometric object around the training data. However, in contrast to SVDD, DMND finds the hyper-ellipsoid of the smallest volume covering the target objects that can contrain outliers by maximizing the determinant of an information matrix. Simulation results are presented for the case when training data are contaminated by compactly located outliers.