THE EFFECTS OF DIFFERENT FEATURE SETS ON THE WEB PAGE CATEGORIZATION PROBLEM USING THE ITERATIVE CROSS-TRAINING ALGORITHM

摘要

The paper presents the effects of different feature sets on the Web page categorization problem. These features are words appearing in the content of a Web page, words appearing on the hyperlinks, which link to the page and words appearing on every headings in the page. The experiments are conducted using a new algorithm called the Iterative Cross-Training algorithm (ICT) which was successfully applied to Thai Web page identification. The main concept of ICT is to iteratively train two sub-classifiers by using unlabeled examples in crossing manner. We compare ICT against supervised naieve Bayes classifier and Co-Training classifier. The experimental results show that ICT obtains the highest performance and the heading feature is considerably succeed in helping classifiers to build the correct model used in the Web page categorization task.