Email has become ubiquitous as a fast and inexpensive form of communication. Algorithms to detect and filter unsolicited email (spam) continue to evolve to maintain their robustness in the face of evermore-sophisticated spammers. This paper evaluates the effectiveness of two different feature selection algorithms on the performance of four different classifiers for spam email detection.
The paper is well organized, but unfortunately many critical explanations are missing or unclear. There is no clear explanation of the fitness function for genetic search, and the greedy algorithm is not well explained. In addition, accuracies are computed but not put in the context of other research results. The authors also report false match rates for their own experiments, but not for previous researchers, weakening their conclusions. A proper hypothesis test would have involved running a Monte Carlo set of experiments to evaluate multiple feature set selections for each selection method. In this way, the average performance of each selection method could be computed for each of the classifiers, providing a stronger statistical basis for the conclusions.
Although this work is well intentioned, it is not clear what actual value it has for researchers in the fields of spam classifiers or feature selection, aside from a brief list of literature on spam detection. Feature subset selection is known to be nondeterministic polynomial-time (NP)-hard, and efficient approximation algorithms are generally application dependent. The authors have provided no insight into why one subset selection algorithm would be superior to the other for the given application. Given the limited level of detail, it is difficult to have confidence in the conclusions.