Multiclass Support Vector Data Description in Extreme Learning Machine Feature Space

Abstract

Abstract Support vector data description (SVDD) method aims to address the one-class classification (OCC) problem to find a hypersphere-shaped description of target data set. For extending SVDD to multiclass classification while remaining the ability of detecting outliers, we propose a novel multiclass SVDD scheme which can be used in effective feature mapping and meta-class separation based on the extreme learning machine algorithm (ELM-MSVDD). Accordingly, the imprecise data difficult to distinguish in specific classes is classified to a meta-class,the meta-class is defined by the disjunction of these specific classes, this operation can reduce the error rate effectively. Experimental results of our ELM-MSVDD method show well performance on the datasets from the UCI machine learning library and radar signal source recognition. Meanwhile, our proposed method provide a theoretical and practical support for other relevant pattern recognition field.