基于小样本学习的垃圾邮件过滤方法

Spam Filtering Method Based on Learning from Small Samples

针对客户端垃圾邮件过滤器难以获取足够训练样本的问题,提出一种基于小样本学习的垃圾邮件过滤方法,利用容易获取的未标记样本提高垃圾邮件过滤的性能。该方法使用已标记的小样本邮件实例集训练一个初始Na?veBayes分类器,以此标注未标记邮件,再使用所有数据训练新的分类器,利用EM算法进行迭代直至收敛。实验结果证明,当给定5个～20个已标记小样本训练邮件时,该方法可有效提高垃圾邮件过滤性能。

It is difficult to collect sufficient labeled E-mails for training a client spam classifier. Aiming at the problem, this paper proposes a spam filtering method based on learning from small samples, which improves the filtering performance with unlabeled samples. An initial Naive Bayes（NB） classifier is trained with a dataset of labeled E-mails, and unlabeled E-mails are probabilistically labeled with it. A new classifier is trained with all E-mails, and iterates to convergence with EM algorithm. Experimental results prove that, given labeled small training samples with a size of 5 to 20, the performance of spam filtering can be effectively improved.