Searching for Interacting Features for Spam Filtering

In this paper, we propose a novel feature selection method— INTERACT to select relevant words of emails for spam email filtering, i.e. classifying an email as spam or legitimate. Four traditional feature selection methods in text categorization domain, Information Gain, Gain Ratio, Chi Squared, and ReliefF, are also used for performance comparison. Three classifiers, Support Vector Machine (SVM), Naïve Bayes and a novel classifier—Locally Weighted learning with Naïve Bayes (LWNB) are discussed in this paper. Four popular datasets are employed as the benchmark corpora in our experiments to examine the capabilities of these five feature selection methods and the three classifiers. In our simulations, we discover that the LWNB improves the Naïve Bayes and gain higher prediction results by learning local models, and its performance is sometimes better than that of the SVM. Our study also shows the INTERACT can result in better performances of classifiers than the other four traditional methods for the spam email filtering.