Combining labeled and unlabeled data with co-training：（与co-training结合标记和未标记数据）.pdfVIP

Combining Lab eled and Unlab eled Data with CoTraining y Avrim Blum Tom Mitchell School of Computer Science School of Computer Science Carnegie Mellon University Carnegie Mellon University Pittsburgh PA Pittsburgh PA avrimcscmuedu mitchellcscmuedu Abstract sults on learning with lopsided misclassica tion noise which we believe may be of inde We consider the problemof using a large unla pendent interest beled sample to boost performance of a learn ing algorithm when only a small set of labeled examples is available In particular we con INTRODUCTION sider a setting in which the description of each example can be partitioned into two distinct In many machine learning settings unlabeled examples views motivated by the task of learning to are signicantly easier to come by than labeled ones classify web pages For example the descrip One example of this is webpage classication tion of a web page can be partitioned into the Suppose that we want a program to electronically visit words occurring on that page and the words somewebsiteanddownloadallthe webpagesofinterest occurringinhyperlinksthatpointtothatpage to us such as all the CS faculty member pages or all We assume that either view of the example the course home pages at some university To train wouldbesucientforlearningifwehadenough