Enhanced One-class Classifiers

As a concept learning method，the one-class classifier is intended to describe the target data,which is different from the two-class ones, thus becoming a new branch of the classifier design andgetting more and more attentions. In this article, some typical research results are overviewed on theone-class classifiers involving density estimation and supporting domain methods. Based on that, ourresearch work is carried out: on one hand, improve a hyperplane model of support-domain method toexploit the prior knowledge of training samples; on the other, extend the design ideas of maximumcontrast classifier to the one-class classifier design.1. Reviewed the key algorithms about one-class classifiers with the viewpoint of the densityestimation and supporting domain, and summarized the evaluation criterion of these methods.Summarized the main methods of one-class classifiers based on density estimation and supportingdomain, further analyzed the advantage and insufficiency of these algorithms. The results are helpfulto make clear the current status of one-class classifier design and the tasks and direction of furtherstudy.2. Mined the prior knowledge of existed normality to improve the generalization capability,designed a new structured one-class support vector machine with local density embedding(ldSOCSVM). Through embedding local information of target data into the structured one-classsupport vector machine, ldSOCSVM not only focuses on the global distribution of normality, but alsotakes the local density information into account, therefore it improves discrimination capability withsparse and robust solution. The strategy proposed above also can be generalized as a unifiedframework applied to most of the existed algorithms.3. Extended the design ideas of maximum contrast classifier to the one-class classifier, designedthe maximum constrained density based one-class classifier (MCDOCC). Inspired by the MaximumContrast Classifiers, MCDOCC solves the classification task within the density-based frameworkwithout giving up the task-oriented specialization of the supporting domain approach and tailored thedensity estimation to the task at hand. It introduces a constrained density by modifying kernel densityestimators with variable mixing weights which are solved by maximizing the average constraineddensity of target samples. By linear programming for optimization of the mixing weights, MCDOCCachieves good generalization performance with sparse representations. Furthermore, MCDOCC withnegative samples (NMCDOCC) is developed, which mines the prior knowledge of few outliers to improve its generalization ability.