BalanSiNG: Fast and Scalable Generation of Realistic Signed Networks

How can we e ciently generate large-scale signed networks following real-world properties? Due to its rich modeling capability of representing trust relations as positive and negative edges, signed networks have spurred much interests with various applications. Despite its importance, however, existing models for generating signed networks do not correctly re ect properties of real-world signed networks. In this paper, we propose BalanSiNG, a novel, scalable, and fully parallelizable method for generating large-scale signed networks following realistic properties. We identify a self-similar balanced structure observed from a real-world signed network, and simulate the self-similarity via Kronecker product. Then, we exploit noise and careful weighting of signs such that our resulting network obeys various properties of real-world signed networks. BalanSiNG is easily parallelizable, and we implement it using Spark. Extensive experiments show that BalanSiNG efciently generates the most realistic signed networks satisfying various desired properties.