A word sequence kernels used in spam-filtering

Abstract

Support vector machines have been widely used in the field of spam filtering, however the majority of used kernels are generally distance-based kernels. Those kernels neglect the structure of the text, and lead to a mass of semantic information lost. This paper proposes the use of word sequence kernel which can mine the information of text structure fully for spam filtering, and propose a word sequence kernel based on dependence measure (DPWSK), which integrates the distinguishing ability of words as prior knowledge into the calculation of kernel functions. DPWSK can reflect the semantic similarity of the text better than WSK and enhance the precision of the SVM filter. The experimental results show that the proposed method can improve the accuracy. Key words: Support vector machines, word sequence kernels, dependence measure, spam filtering.