摘要
在文本分类中获得有类别标记训练样本的代价是很高昂的,本文针对这个问题对传统的模糊聚类方法进行改进,提出模糊划分聚类方法 FPCM,将聚类的无监督性和样本的先验知识结合起来,通过相似度度量聚类相关文本,取得比较客观的簇和少量标记文本,为监督学习找到分类依据,并结合朴素贝叶斯增量学习方式进行分类器的学习。本文进一步用估计分类误差损失的方法平衡选取候选样本,提高了分类准确率,实现了应用范围更加广泛的无标记文本分类学习模型。
Bayes learning theory is to obtain estimate of non-labeled samples by transcendental information and sample data. The application of text classification is to classify non-labeled texts by learning labeled class samples. But it is very difficult to obtain labeled training samples. In the paper the problem is analyzed in point of clustering view. The clustering is a non-supervised learning method, and has a character of independence on defined classes and labeled training samples. The thesis improve on tradition fuzzy clustering to bring forward Fuzzy Partition Clustering Method （FPCM）. FPCM is a dynamic clustering method based on centroid technique. A few labeled texts are obtained to find classification foundation for supervised learning by fuzzy Partition clustering non-labeled Web texts. The sample＇ s transcendental knowledge and clustering＇s non-supervisory are combined, and correlation texts are clustered by measuring similar degree. Naive Bayes augment learning style is further used to design and learn classifier. At the same time, classification precision is advanced using the way of selecting balance candidate samples after estimating the loss of classifying error. The model of text classifying using non-labeled training sample with more extensive application is realized.
出处
《计算机科学》
CSCD
北大核心
2006年第3期200-201,211,共3页
Computer Science
基金
973国家重点基础研究项目(G1998030414)
北京市优秀人才专项经费资助项目(20042D0501604)
