Uninterrupted Approaches for Spam Detection Based on SVM and AIS

The proliferation of spam has increasingly caused serious problems to our daily electronic communications. Jupiter Research estimates an e-mail user will on average receive more than 3,900 spam mails in 2007. This number is on average of 40 spam mails in 1999. This paper proposes uninterrupted detection approaches based on Incremental Support Vector Machine and Artiϐicial Immune System for the spam of e-mail stream. These approaches use a window to hold several classiϐiers, each one classiϐies the e-mail independently, and the e-mail is labeled according to a majority voting strategy. The exceeding margin update technique of support vector machine (SVM) is also used for the dynamic update of each classiϐier in the window. A sliding window is used to purge out-of-date knowledge. When a new batch of e-mail arrives, the classiϐier at rightmost in the window is removed from the window while the remaining classiϐiers just slide a position to right and the classiϐier at leftmost is newly generated by the previous batch. These two techniques endow our algorithms with dynamic and adaptive properties as well as the ability to trace the changing content of e-mails and user's interests in an uninterrupted way. Eight methods, possibly regarded as different implementations of the uninterrupted detection of e-mail stream, including Hamming Distance (with and without mutation), Included Angle, SVM and Weighted Voting, are developed and elaborated in this paper. Experiments on two public benchmark corpora PU1 and Ling are conducted to verify the validity of the proposed methods. The eight methods are compared with current methods for accuracy, precision, recall, miss rate and speed of detection. The results demonstrate that the proposed uninterrupted detection approaches are promising way to contain spam.