CAS-ICT at TREC 2005 SPAM Track: Using Non-Textual Information to Improve Spam Filtering Performance

This paper introduces our work in the TREC2005 SPAM track. Naïve Bayes and Littlestone’s Winnow are chosen as our basic classifiers. In our investigation, we found that when the structures of Ham and Spam are very different, the feature distributions of them vary a lot. Thus the factor of structure is introduced into our filter. Besides textual word feature, some kind of other features are also considered in our filter. Our experimental results show that Winnow outperforms Naïve Bayes and the multi-feature model outperforms structure based model.