MailMind A Connectionist E-Mail Sorting Agent

MailMind is an application of backpropagation networks to the real world problem of sorting e-mail messages based upon the sender’s address. This paper presents the strategies employed in the project to adapt to the greatly varying properties of datasets in this domain, such as scale and time dependent change of the problem as well as extremes in noisiness and lack of variation. These strategies include a number of encoding algorithms for e-mail addresses, which can speed up the learning task and support generalization, and an object oriented implementation approach, which supports dynamical changes to the network structure. One possible concretization of such changes is given through an algorithm that utilizes the object oriented structure to extract subnets and create a network hierarchy. Included empirical results reveal the success of the employed strategies in solving the posed problem.