Optimising Näıve Bayesian Networks for Spam Detection

In 2001, Spam e-mails accounted for 8% of all e-mails sent over the internet. By 2002, that number had risen to 36%. Despite these staggering numbers, there are only about 150 people responsible for the bulk of spam e-mail. Because of this, there are many words and phrases common to most spam e-mails that do not occur in desirable e-mail messages. A näıve Bayesian classifier can be employed to detect these spam messages, but the computing costs associated with classifying a large number of messages can be very high. To resolve this problem, two modifications are made to the classifier. The computation is simplified using unique classifiers for each message. In addition, the vocabulary used by the classifiers is reduced using an entropy-based technique. These modifications reduce the cost of the classification and reduce the error rate of the classifier.