A New E-mail Agent Architecture Based on Semi-supervised Bayesian Networks

A new e-mail agent architecture with a Bayesian network (BN) has been investigated in order to detect important e-mail (IM) of office users. The BN has nodes related with users´ resultant behaviors during the e-mail operation, which enables to adapt the agent for users´ intentions by implicit feedbacks, called semi-supervised learning. We have investigated 5 examinees for 2 months. It is certain that our BN is so effective for the detection of IM, because we obtain an accuracy of as high as 0.924 by a fully supervised learning. Moreover, the similar accuracy can be obtained in the semi-supervised learning, where the nodes of resultant behaviors can be properly working in a cycle of the implicit feedback, as an alternative for users´ questionnaires