Transfer Learning for Enhancing Information Flow in Organizations and Social Networks

The task of suggesting recipients for an email has recently received attention as it has potential to enhance the flow of knowledge and information within an organization or social network. We investigate two transfer learning techniques to improve recipient prediction performance through considering predictions for multiple users. We present a novel continuous hidden variable conditional random field for the recipient prediction problem. We characterize this construction as a type of discriminative author recipient topic or DART model. First we show transfer based performance increases achieved through shared hidden variables for prediction across different users. Second, we show how transfer from an organization wide model to a user specific model through parameter prior structure also confers substantial advantage, especially when models are constructed for new users.