【作者】 许震； 沙朝锋； 王晓玲； 周傲英；

【Author】 Xu Zhen~1,Sha Chaofeng~1,Wang Xiaoling~2 and Zhou Aoying~(2,3) 1(School of Computer Science,Fudan University,Shanghai 200433) 2(Institute of Massive Computing,Software Engineering Institute,East China Normal University,Shanghai 200062) 3(Shanghai Key Laboratory of Intelligent Information Processing,Fudan University,Shanghai 200433)

【机构】 复旦大学计算机科学技术学院； 华东师范大学海量计算研究所； 上海市智能信息处理重点实验室；

【摘要】 在实际应用中,由于各种原因,时常无法直接获得反例,导致传统分类方法的暂时失灵,所以,基于正例和未标识集的主动分类问题顿时成为了理论界的焦点。为了解决这一难题,理论界相继提出了许多不同的解决方案,然而,这些方法不能有效处理非平衡问题,尤其当隐匿反例非常少或训练集中的实例分布不均匀时。因此,本文提出了一种基于KL距离的主动分类算法—LiPU:依次挖掘出未标识集中的最可靠正例和反例,接着使用训练好的增强型分类器来找出所有反例。与其他方法相比,不仅提高了分类的查准率和查全率,而且具有鲁棒性。更多还原

【Abstract】 In many real applications,it’s often very difficult or expensive to get labeled negative examples,thus the traditional classification techniques can’t work effectively.As a result,the active classification from positive and unlabeled data grows important,getting much focus.In recent years,lots of methods are proposed,but most of them can’t always work well,especially when the number of negative examples hidden in the unlabeled set is very small or the size of positive sub-classes in the training set differs.In this paper,a novel KL divergence-based classification algorithm,named LiPU,is proposed to tackle this special problem.It firstly finds reliable positive and negative examples existing in the unlabeled set,followed by an enhanced logistic regression classifier to classify the unlabeled set.The experiments show that the proposed approach not only improves precision and recall,but also robust,compared with former work in the literature.更多还原