Spam filtering is a principle anti-spam technique in the tag-and-tug war between spams and anti-spams. Most spam filtering techniques, such as SVM, K-NN, Boosting, Winnow, and Bayesian filtering, are based on machine learning and content analysis. The problems of these techniques are the unsatisfactory recall rates, long training time, and high false alarm rates.This paper proposes a three-dimensional hybrid spam filtering scheme consisted of the following three filtering technologies: collaborative spam filtering based on user feedbacks, whitelist spam filtering based on personal email network, and adaptive Bayesian filtering.Collaborative spam filtering targets mass spam mailing. It uses a modified Nilsimsa abstract algorithm to differentiate similar spams and at the same time makes use of direct and indirect feedback collecting techniques. Whitelist spam filtering targets mass legitimate mailing and works by calculating the clustering coefficient of personal email networks. Adaptive Bayesian filtering builds on top of the previous two techniques and makes use of their outputs in its multi-iteration training.Experiments show that our system improves the recall rate by 4.26%, and the precision rate by 0.27% compared with Naive Bayesian Filtering. It reduces user's Total Cost Ratio by 15%. |