Research On One-class Classifier Based On Geometric Covering Model Of Target Class In High-dimensional Space

Traditional pattern recognition tasks try to distinguish between two or more classes with the training set containing objects from all the classes. However, the need to train classification models from one-class data alone arises in many applications. In such problems, only one-class of the data, called the target set, are available, so one-class classifiers should learn from the target class and form the data covering model for classification. Up to now, the research of one-class classification has attracted much attention. However, the one-class classification of irregular and complex data in high-dimensional space is still a difficult problem. Based on the achievements of previous researchers, this paper conducts the research on some novel one-class classification covering models in high-dimensional.Firstly, a sparse distance metric learning algorithm in high-dimensional space is proposed to improve the descriptive performance of one-class classifiers in this paper. The learned metric can be easily embedded into one-class classifiers, and effectively improve the descriptive performance of one-class classifiers. It makes a stronger generalization ability of one-class classifiers.Secondly, a one-class classification algorithm based on sparse minimum spanning tree (SMST) covering model is presented. The method firsly constructs sparse k-nearest-neighbor graph representation for the target class; then a recursive graph bipartition algorithm is introduced to find the micro-cluster; finally it builds minimum spanning tree on the centers of micro-cluster. Experimental results demonstrate the effectiveness of the method.Thirdly, Minimum Spanning Tree Class Descriptor (MSTCD) describes the target class with too many branches, and its local coverage is not so reasonable. In this case, a one-class classification algorithm based on Steiner minimal tree of typical samples (TS-SMT) covering model is presented in this paper. The method firstly prunes the training set. Then it builds Steiner minimal tree covering model on the retained typical samples. Experimental results show good performance of this method.Finally, in order to construct the tight covering model of target class, the convex hull data description (CHDD) based one-class classifier is presented in this paper. The model is a non-parametric classifier which covers the irregular data adaptively in the feature spaces. By the introduction of kernel functions, the stronger ability of nonlinear classification can be obtained. Experimental results show good performance of the new classifier.