Learning to Filter Junk E-Mail from Positive and Unlabeled Examples

摘要

We study the applicability of partially supervised text classification to junk mail filtering, where a given set of junk messages serve as positive examples while the messages received by a user are unlabeled examples, but there are no negative examples. Supplying a junk mail filter with a large set of junk mails could result in an algorithm that learns to filter junk mail without user intervention and thus would significantly improve the usability of an e-mail client. We study several learning algorithms that take care of the unlabeled examples in different ways and present experimental results.