Genetic-based feature selection for spam detection

In recent years, email has evolved into a pervasive and economical means of communication, but spam as an annoying fact has decreased usefulness of this means. For encountering the challenge, email filtering as a special kind of text classification emerged and developed. A main problem in text classification tasks which is more serious in email filtering is existence of large number of features. For solving the issue, various feature selection methods are considered, which extract a lower dimensional feature space from original one and offer it as input to classifier. In this regard, we examined effectiveness of two existent individual methods and offer a new combinational method. The methods, which are experimented individually, are Information Gain (IG) and χ2 statistic (CHI), and our combined method is applying genetic algorithm (GA) on the top features selected by IG. We used Perceptron neural network as classifier. For evaluation of our system, experiments were conducted on PU data set. The results showed that the individual methods are very effective in reducing dimensionality of input space along with increasing performance of classifier, and the combined method further improves performance in spite of bringing dimensionality to a lower extent.