Abstract
Characterizing and modeling the distribution of a particular family of graphs are essential for the studying real-world networks in a broad spectrum of disciplines, ranging from market-basket analysis to biology, from social science to neuroscience. However, it is unclear how to model these complex graph organizations and learn generative models from an observed graph. The key challenges stem from the non-unique, high-dimensional nature of graphs, as well as graph community structures at different granularity levels. In this paper, we propose a multi-scale graph generative model named Misc-GAN, which models the underlying distribution of graph structures at different levels of granularity, and then "transfers" such hierarchical distribution from the graphs in the domain of interest, to a unique graph representation. The empirical results on seven real data sets demonstrate the effectiveness of the proposed framework.