We add a new phase, called reforming phase, to support vector data description (SVDD) between the training and testing phases. The reforming phase enables us to reconsider the SVDD’s assumption of the uniformity of features in calculating the distance of an object to the center of hypersphere. In the reforming phase, the features are assumed as a group of experts who have different impacts in overall outlier detection. In doing so, the proportion of each feature in the distance of an object to the center of hypersphere is specified. Subsequently, the opinions of the experts about the label of the corresponding object are determined based on these measured proportions. By using different group decision-making methods for aggregating the opinions of the experts, a large variety of new models are obtained based on one SVDD’s trained model. Specially, we utilize a kind of ordered weighted averaging operator as group decision-making method and introduce cDFS-SVDD based on this method. cDFS-SVDD performs runtime feature selection and calculates the distance of an object to the center of hypersphere dynamically at test time based on these selected features. We apply the method to the anomaly detection problem in mobile ad hoc networks as well as two UCI datasets by which the performance of SVDD improves significantly in separating the target and outlier objects.
Rahmanimanesh, Mohammad, Jalal A. Nasiri, Saeid Jalili, and Nasrolah Moghaddam Charkari. 2019. Adaptive three‑phase support vector data description. Pattern Analysis and Applications ۲۲ (۲): ۴۹۱-۵۰۴.