USpam -- A User Centric Ontology Driven Spam Detection System

Recently, content-based Spam detection frameworks are receiving a significant amount of attention by academic researchers and industrial practitioners. However, the anticipated wide scale proliferation is limited (mainly) because of two important shortcomings: (1) high false alarm rate that results in moving legitimate messages into Spam folders, and (2) inability to self-learn a user's profile, as a result, they are unable to identify useful Spam (we call it Good Spam) that might be of great interest to a user's personal or business aspirations. In this paper, we propose USpam, a system that uses ontologies to model features that are extracted rom a user's profile. The features are given to machine learning classifiers J48 and Naive Bayes -- that learn a user centric model of Good Spam or Bad Spam. As a result, the system puts a message into a user's inbox if its contents are relevant to his interests. The USpam is evaluated on NRON Spam datasets, and the results of experiments reveal that false alarms are reduced by 10% to 30% compared with existing prior art without compromising the detection accuracy.