Modeling Sequential Representation with Interaction Dynamics for Social Prediction

Sequential prediction is of great value for resource allocation in our society, such as recommendation and urban planning. To train an effective sequential prediction model, it is a common practice to optimize the representation learning of historical item-item transitions. However, real-world sequential behaviors are constantly affected by dynamic environments, which makes the training network suffer from model uncertainty. It is a challenge task to incorporate such dynamic uncertainty into the learning of sequential patterns. In this paper, we rethink the dynamics in interaction sequences, and present a novel sequential representation learning framework DSeq which models dynamics from three levels: network architecture, training strategy and uncertainty estimation.