Spam Filtering using Local-global Bayesian Classifier

Abstract

Spam is an email, which is usually sent in bulk by the sender. Unlike legitimate mails, there is no agreement between the receiver and the sender of the mail. That's why they are also termed as unsolicited mails. To prevent the delivery of this so called spam messages, an automated tool called a spam filter is used to recognize spam. As there is no single definition of spam, it is difficult to formulate rules to block such unwanted messages. There are several techniques used to stop those unwanted messages. It is not full proof against spam, even with the introduction of new state of the art techniques. Some of the techniques are based on manually configured rules, others rely on statistical calculations for adapting themselves according to the current situation. In this thesis, a novel learning framework for classification of messages into spam and legit is proposed. Naive Bayes (NB) model is a statistical filtering process which uses previously gathered knowledge. Instead of using a single classifier, the use of local and global classifier, based on the Bayesian hierarchal framework is proposed. This helps in achieving multi-task learning, as simultaneous extraction of knowledge can be achieved while achieving classification accuracy. Knowledge among different task can be shared while learning for task specific.