GSCALER: Synthetically Scaling A Given Graph

Enterprises and researchers often have datasets that can be represented as graphs (e.g. social networks). The owner of a large graph may want to scale it down to a smaller version, e.g. for application development. On the other hand, the owner of a small graph may want to scale it up to a larger version, e.g. to test system scalability. This paper investigates the Graph Scaling Problem (GSP): Given a directed graph G and positive integers ñ and m̃, generate a similar directed graph G̃ with ñ nodes and m̃ edges. This paper presents a graph scaling algorithm Gscaler for GSP. Analogous to DNA shotgun sequencing, Gscaler, decomposes G into small pieces, scales them, then uses the scaled pieces to construct G̃. This construction is based on the indegree/outdegree correlation of nodes and edges. Extensive tests with real graphs show that Gscaler is scalable and, for many graph properties, it generates a G̃ that has greater similarity to G than other state-of-the-art solutions, like Stochastic Kronecker Graph and UpSizeR.