An adaptive concentration selection model for spam detection

Abstract

Concentration based feature construction (CFC) approach has been proposed for spam detection. In the CFC approach, Global concentration (GC) and local concentration (LC) are used independently to convert emails to 2-dimensional or 2n-dimensional feature vectors. In this paper, we propose a novel model which selects concentration construction methods adaptively according to the match between testing samples and different kinds of concentration features. By determining which concentration construction method is proper for the current sample, the email is transformed into a corresponding concentration feature vector, which will be further employed by classification techniques in order to obtain the corresponding class. The k-nearest neighbor method is introduced in experiments to evaluate the proposed concentration selection model on the classic and standard corpora, namely PU1, PU2, PU3 and PUA. Experimental results demonstrate that the model performs better than using GC or LC separately, which provides support to the effectiveness of the proposed model and endows it with application in the real world.