摘要： 提出一种在选取特征时考虑特征与特征之间联系的算法。对特征词之间的关联关系进行挖掘，找出那些对类别有重要影响的特征词组，而这些特征词组中的每个单词在传统单独打分策略的特征选择算法中很可能会因分值过低而被丢弃。在Ruters21578、20Newsgroup文本数据集上进行实验，将算法与广泛应用的特征选择算法(信息增益、CHI等)进行对比、分析。实验结果表明该方法是一种有特点、有效的特征选择方法。

关键词: 特征选择, 交叉熵, 文本分类, 关联挖掘

Abstract: This paper proposes a method, which considers the relationship between two words in feature selection. The relationship between two words which have significant impact on classification is mined, and two-word-sets are found out. Some words in these sets may be discarded due to low scores achieved through the conventional feature selection methods. The algorithm is compared with other conventional feature selection approaches: Information Gain(IG), CHI, etc. Experimental results on Ruters21578 dataset and 20Newsgroup dataset prove that the proposed method is effective to others.

Key words: feature selection, cross-entropy, text classification, association mining

中图分类号: 

• TP391