A Clustering Based Feature Selection Method in Spam Detection

One of the ways to detect spam is classifying emails into two categories: spam and non-spam. The high efficiency of machine learning methods in various fields has developed them in text clasification problems. The mechanism of machine learning-based classifiers that classify emails according to their content is based on a set of features, where due to the high volume of emails, using an efficient feature reduction algorithm plays an important role. Unlike the previous methods which select only the superior features and ignore the rest of the unselected features, in the proposed method of this article we try to use unselected features as well. The method is that after applying an initial feature selection, the unselected features are clustered and then each cluster is mapped to a new feature and the final feature vector forms from the selected ones and those mapped from the clusters. In this study, by applying two methods of selecting the initial feature and also two mapping functions, four methods were presented and analyzed using two datasets PU2 and PU3. The results of the analysis showed that the method based on feature selection DF and the advanced mapping function has the highest efficiency among all the proposed methods. Also, the proposed methods are more efficient than base feature selection methods (without clustering).