Inferring Diffusion Networks with Sparse Cascades by Structure Transfer

Inferring diffusion networks from traces of cascades has been intensively studied to gain a better understanding of information diffusion. Traditional methods normally formulate a generative model to find the network that can generate the cascades with the maximum likelihood. The performance of such methods largely depends on sufficient cascades spreading in the network. In many real-world scenarios, however, the cascades may be rare. The very sparse data make accurately inferring the diffusion network extremely challenging. To address this issue, in this paper we study the problem of transferring structure knowledge from an external diffusion network with sufficient cascade data to help infer the hidden diffusion network with sparse cascades. To this end, we first consider the network inference problem from a new angle: link prediction. This transformation enables us to apply transfer learning techniques to predict the hidden links with the help of a large volume of cascades and observed links in the external network. Meanwhile, to integrate the structure and cascade knowledge of the two networks, we propose a unified optimization framework TrNetInf. We conduct extensive experiments on two real-world datasets: MemeTracker and Aminer. The results demonstrate the effectiveness of the proposed TrNetInf in addressing the network inference problem with insufficient cascades.