A Preliminaries

We now give further description of some background material. In order to evaluate the quality of the sampled graphs (i.e., to decide whether the model One robust way to achieve this is to define the metric on latent vector representation spaces, and expect representations of graphs rather than the original objects. The estimator, however, has low variance but high bias [4]. PR is also shown to correlate with human judgments in visual domain. DC is shown to be more robust compared to PR. Maximum Mean Discrepancy (MMD) [15] compares two distributions of any type, based on item-level comparison by a kernel function (e.g., polynomial kernel K (x We use naming from this figure for the proof.