基于正例和未标文档的半监督分类研究  

Research on Semi-Supervised Classification Based on Positive and Unlabeled Text Document

作　　者：范新[1] 沈闻[1] 丁泉勋[1] 沈洁[1] 

机构地区：[1]扬州大学信息工程学院计算机科学系,江苏扬州225009

出　　处：《计算机技术与发展》2009年第6期58-60,64,共4页Computer Technology and Development

基　　金：国家自然科学基金资助项目(60673060)

摘　　要：针对负例类别很难获得训练样本的情况,提出了一种基于正例和未标文档的半监督分类方法。已知仅有正例文本的情况下,引入k-means聚类算法对未标样本集进行聚类,从未标样本集中选出最为可靠的负例样本信息,初始化分类器。基于EM的极大似然估计理论,在每步迭代的E-step中,将中间分类器最有把握对其类别进行预测的未标注样本进行分类,并应用到M-step中修正分类器的参数值,迭代选择最优分类器。实验结果表明,该方法取得了较好的分类效果。Presents a high performance method classifying positive and unlabeled documents. The idea is to first use cluster to extract same reliable negative documents from the unlabeled set and initial a classifier. Then optimize our classifier with the expectation - maximization （EM） algorithm. In each E - step,would like to annotate the most reliable documents, which processed in M - step. After several iteration,ean select a better classifier with EM algorithm. The experiments show that this method achieves a high performance.

关 键 词：聚类 K-MEANS EM 分类 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]