ForGen: Autoregressive Generation of Sparse Graphs with Preferential Forest.

Graph is a natural way to model interactions between objects, such as in the field of biology and social science. Over the past decades, modeling and generating graphs have been a popular research topic, largely inspired by observed properties of real-world graphs. Since traditional approaches relying on hand-crafted mechanisms are only capable of capturing some specific graph properties, recent focus has been shifted to deep neural methods. However, the task is still challenging in terms of efficiency. To address this issue, we observe that the connectivity (i.e., degree) of nodes follows scale-free distribution for most real-world graphs, which can be utilized to accelerate the generation process. We propose ForGen, a Forest-based Generation model that contains a graph-level and an edge-level autoregressive generator. Specifically, for the edge-level model, motivated by the skewed distribution of node degree and the Huffman tree, we design a forest-like data structure to accelerate edge connection via shallow tree searches and better parallelism. Experiments on both synthetic and real-world graph datasets show that ForGen is two times faster than the current state-of-the-art method for graph generation, and guarantees better generated graph quality.