Performance evaluation of classifiers for spam detection with benchmark datasets

Detection of unwanted, unsolicited mails called spam from email is an interesting area of research. Researchers with the help of machine learning algorithms normally find the best classifier that distinguishes a spam from a benign mail called ham. It is necessary to evaluate the performance of any new spam classifier using standard data sets. The public corpora of email data sets that are available has certain special characteristics that reflects the time of compilation, the number of users considered and the general subject of the messages. This paper describes a comprehensive study on the performance evaluation of various machine learning algorithms using two benchmark data sets. The evaluations clearly demonstrate the superior performance of the tree classifiers and ensemble based classifiers with trees as basic classifier. Both the tree classifier and the ensemble classifier were performing with accuracy greater than 96% and mean absolute error less than 0.05%.