关键词: 文本分类, 过滤 , 噪音特征项, 层次特征选择, hRIPPER, RIPPER

Abstract: The rule-based text categorization algorithm RIPPER was specialized with easy understanding, quick optimization, and high efficiency. However, when the rule refers to too many features, not only were the above advantages apparently weakened, but also the performance of the algorithm decreases. The hierarchy-based hRIPPER though uses hierarchical feature selection and can still not filter features fully. Then an improved text categorization algorithm iRIPPER was proposed to solve the problems in the learning process of RIPPER and hRIPPER, which filters features more thoroughly during the learning process. The experiment proves that it selects features effectively, generates fewer rules, and reduces the time in the growing process. Therefore it improves the performance of the rule-based text categorization.

Key words: filter , noisy feature set, hierarchical feature selection, hRIPPER, RIPPER, text categorization