Continuous Time Information Network Mining for Diffusion Cascades

Information network mining has various applications in viral marketing, link prediction, recommendation and many other aspects. Traditionally, information network mining assumes a discrete time step in activity logs, thus is not capable to do continuous time predictions. Secondly, most information network mining models and algorithms only consider pairwise influence. According to our observation, indirect influence from τ neighbours (τ > 1), and structural influence also contribute to information propagation. In this paper, we formally define the indirect influence and structural influence and consider these two effects when doing the network mining in our model ConInfMine. Furthermore, we use the network backbone as a input of our modelling. We perform some theoretical analysis on the convexity of the problem and convergence of the training procedure. Finally, experiments on both synthetic and real-world datasets demonstrates the effectiveness of our algorithm.