Motif-aware diffusion network inference

Diffusion processes over networks are widely encountered in various real-world applications. In many scenarios, however, only the states of nodes can be observed while the underlying diffusion networks are unknown. Many methods have therefore been proposed to infer the underlying networks based on node observations. To enhance the inference performance, structural priors of the networks, such as sparsity, scale-free, and community structures, are often incorporated into the learning procedure. As the building blocks of complex networks, motifs occur frequently in many real-world networks and play a vital role in describing the network structures and functionalities. However, to the best of our knowledge, no existing work has attempted to infer the diffusion network through exploiting network motifs. In order to explore this uncharted yet important issue, in this paper, we propose a novel framework called motif-aware diffusion network inference (MADNI), which aims to mine the motif profile from the node observations and infer the underlying network based on the mined motif profile. The mined motif profile and the inferred network are alternately refined until the learning procedure converges. We derive two network inference models in our framework for inferring the individual-level diffusion network and the meta-population-level diffusion network, respectively. Extensive experiments on both synthetic and real-world datasets validate the effectiveness of the proposed framework. Moreover, we show the generality of MADNI by analyzing the relations between mesoscale motif prior in our framework and other high-level structural priors such as community and scale-free properties in the existing works.