Feature-Enhanced Probabilistic Models for Diffusion Network Inference

Cascading processes, such as disease contagion, viral marketing, and information diffusion, are a pervasive phenomenon in many types of networks. The problem of devising intervention strategies to facilitate or inhibit such processes has recently received considerable attention. However, a major challenge is that the underlying network is often unknown. In this paper, we revisit the problem of inferring latent network structure given observations from a diffusion process, such as the spread of trending topics in social media. We define a family of novel probabilistic models that can explain recurrent cascading behavior, and take into account not only the time differences between events but also a richer set of additional features. We show that MAP inference is tractable and can therefore scale to very large real-world networks. Further, we demonstrate the effectiveness of our approach by inferring the underlying network structure of a subset of the popular Twitter following network by analyzing the topics of a large number of messages posted by users over a 10-month period. Experimental results show that our models accurately recover the links of the Twitter network, and significantly improve the performance over previous models based entirely on time.