作　　者： (董勇）; (孙广玲）; (刘志）;

机构地区： 上海大学通信与信息工程学院,上海200444

出　　处： 《上海大学学报（自然科学版）》 2017年第4期524-534,共11页

摘　　要： 在机器学习中,当测试阶段无法得到训练阶段拥有的特权信息时,特权学习(learning using privileged information,LUPI)是一个有效的解决框架.由于获取特权信息需要特殊的条件,或由于其他原因,往往不能获得全部训练样本的特权信息,因此提出了一种直观却有效的方法.对于缺失特权信息的这部分训练样本,将它们的可用信息同时用作特权信息,并将其纳入到支持向量机(support vector machine+,SVM+)的模型中,引入了一种新的扩展SVM+(extended SVM+,e SVM+)模型.进一步地,对于不涉及特权信息的常规有监督学习问题,也将训练样本的特征(可用信息)同时用作特权信息,引出一种新的扩展SVM模型(e SVM),e SVM也可认为是SVM+的特例.在两个公开的人脸表情数据库BU-3DFE和Bosphorus上进行了实验,结果证实了将可用信息用作特权信息策略的有效性. In machine learning, some information is only available during learning phase. Learning using privileged information （LUPI） can provide an effective solution to the problem. Such information is called privileged information. To deal with the issue that only partial privileged information of training data is available, this paper presents an intuitive but effective strategy called extended support vector machine＋ （eSVM＋）. Specifically, for data without privileged information, available information is used as privileged information simultaneously and further cooperate it into the SVM＋ formulation. In addition, for a regular supervised learning paradigm, a similar idea is adopted that all training data are both available and privileged. Naturally, it is extended SVM （eSVM）, and also a special aspect of SVM＋. Experimental results show that the proposed strategy can boost generalization performance of the classifier on two benchmark expression databases, BU-3DFE and Bosphorus.

关 键 词： 特权学习 可用信息 特权信息 支持向量机