Spam Detection using Generalized Additive Neural Networks

During the last decade the number of spam messages sent has increased significantly. These undesired emails place a heavy burden on end users and email service providers. As a result, a tenacious struggle to outsmart each other exists between people who send spam and the spam filter providers. Constant innovation is therefore of vital importance to curb the rapid increase of spam. In this article a Generalized Additive Neural Network (GANN) is harnessed to detect spam. This relative new type of neural network has a number of strengths that makes it a suitable classifier of spam. An automated GANN construction algorithm is applied to a spam data set. This method can perform in-sample model selection or cross-validation and produces results that compare favourable to other classifiers found in the literature. Even though neural networks in general are regarded as black box techniques, results obtained by GANNs can be interpreted by graphical methods. Partial residual plots assist the spam researcher to gain insight into the models constructed. Finally, by performing variable selection the temporal evolution of spam can be tracked. This feature ensures that the models can adapt to the ever-changing tactics of people who send spam with greater ease. Index terms Generalized Additive Model, GAM, Generalized Additive Neural Network, GANN, Neural Network, Spam, Spam detection.