Pairwise-interactions-based Bayesian Inference of Network Structure from Information Cascades

An explicit network structure plays an important role when analyzing and understanding diffusion processes. In many scenarios, however, the interactions between nodes in an underlying network are unavailable. Although many methods for inferring a network structure from observed cascades have been proposed, they did not perceive the relationship between pairwise interactions in a cascade. Therefore, this paper proposes a Pairwise-interactions-based Bayesian Inference method (named PBI) to infer the underlying diffusion network structure. More specifically, to get more accurate inference results, we measure the weights of each candidate pairwise interaction in different cascades and add them to the likelihood of a contagion process. In addition, a pre-pruning work is introduced for candidate edges to further improve the inference efficiency. Experiments on synthetic and real-world networks show that PBI achieves significantly better results.