.邮件过滤中特征选择方法的性能评价与分析*

基于内容的邮件过滤本质是二值文本分类问题。特征选择在分类之前约简特征空间以减少分类器在计算和存储上的开销,同时过滤部分噪声以提高分类的准确性,是影响邮件过滤准确性和时效性的重要因素。但各特征选择算法在同一评价环境中性能不同,且对分类器和数据集分布特征具有依赖性。结合邮件过滤自身特点,从分类器适应性、数据集依赖性及时间复杂度三个方面评价与分析各特征选择算法在邮件过滤领域的性能。实验结果表明,优势率和文档频数用于邮件过滤时垃圾邮件识别的准确率较高,运算时间较少。

Evaluation and analysis of feature selection methods for e-mail filtering

The nature of content-based e-mail filtering is a binary text classification problem. Feature selection methods reduced the feature dimension before classifying e-mails in order to reduce the cost of computing and storage, while filtering some noise features to improve the classification accuracy. Feature selection was an important factor which decided the accuracy and timeliness of e-mail filtering. However, every feature selection algorithm had different performance in the same environment, and was affected by classifiers and data distribution. Combining characteristics of e-mail filtering, this paper evaluated and analized the following aspects of feature selection methods which used to filter e-mails: classifier adaptability, data set dependence, time complexity. Experimental results show that odds ratio and document frequency have higher accuracy and less computing time when they are used to filter emails.