The Internet emerged as a powerful infrastructure for the worldwide communication and interaction of people. Some unethical uses of this technology (for instance spam or viruses) generated challenges in the development of mechanisms to guarantee an affordable and secure experience concerning its usage. This study deals with the massive delivery of unwanted content or advertising campaigns without the accordance of target users (also known as spam). Currently, words (tokens) are selected by using feature selection schemes; they are then used to create feature vectors for training different Machine Learning (ML) approaches. This study introduces a new feature selection method able to take advantage of a semantic ontology to group words into topics and use them to build feature vectors.
To this end, we have compared the performance of nine well-known ML approaches in conjunction with (i) Information Gain, the most popular feature selection method in the spam-filtering domain and (ii) Latent Dirichlet Allocation, a generative statistical model that allows sets of observations to be explained by unobserved groups that describe why some parts of the data are similar, and (iii) our semantic-based feature selection proposal. Results have shown the suitability and additional benefits of topic-driven methods to develop and deploy high-performance spam filters.