Web Spamming By Detecting Gray Mail

Improving Email Filtering Performance Using Gray Mail Detection

Abstract

Weaddress the problem of gray mail – messages that could reasonably be consideredeither spam or good. Email users often disagree on this mail, presentingserious challenges to spam filters in both model training and evaluation. Inthis paper, we propose four simple methods for detecting gray mail and comparetheir performance using recall-precision curves. Among them, we found thatemail campaigns that have messages labeled differently are the most reliable source for learning a graymail detector. Preliminary experiments also show that even when the gray maildetector is imperfect, a traditional statistical spam filter can still beimproved consistently in different regions of theROC curve by incorporating this new information.