Application of Linear Classifier on Chinese Spam Filtering

摘要

Spam is a key problem in electronic communication. Especially in large-scale email systems. Content-based filtering is one mainstream method of combating this threat in its forms, an e-mail filtering system can learn directly from a user's mail set, but the previous Content-based filtering methods are hard to find a balance between efficiency and effectiveness. Such algorithms of text categorization as Naive Bayes, kNN, Decision Tree and Boosting can be applied in spam filtering. However, the effectiveness of Naive Bayes is limited and it is not fit for instant feedback learning. Others algorithm such as SVM are more effective but complicated to compute. Because in a real email system a large volume of emails often need to be handled in a short time, efficiency will often be as important as effectiveness when implementing an anti-spam filtering method. So we intend to find a linear classifier to solve this problem, two online linear classifiers: the Perception and Winnow were explored for this task, which are two fast linear classifiers. The training of these two methods is online and mistake driven. Furthermore, they are suitable for feedback. We employ the two methods in three benchmark corpora, including PU1, Ling spam and 2005-Jun, the experiments in public e-mail corpus show an effective result. We conclude that the two online linear classifiers have a state-of-the-art performance for filtering spam, especially for Chinese spam emails.