Local Spectral for Polarized Communities Search in Attributed Signed Network

Fanyi Yang, Huifang Ma, Wentao Wang, Zhixin Li, Liang Chang

Abstract: Signed networks are graphs with edge annotations to indicate whether each interaction is friendly (positive edge) or antagonistic (negative edge). Community search on signed network expects to explore the polarized communities (i.e., two antagonistic subgraphs) containing the set of query nodes. Though previous studies have been proven effective, they generally ignore two insights. First, node attributes provide side information to describe features of nodes. It contributes to optimal results. Secondly, the problem of detecting polarized communities from a global perspective is increasingly limiting and is computationally expensive on large-scale networks. These aspects motivate us to develop a new community search framework searching for Polarized Communities in Attributed Signed network (PCAS). Specifically, we propose a new strategy to combine node attributes with signed topology, which helps to make the most of the different dimensions of information. Furthermore, to search for polarized communities containing the set of query nodes, a sparse indicator-vector is developed based on Rayleigh quotient via solving a linear programming problem. Extensive experimental results on two real-world attributed signed graphs have demonstrated the discovered polarized communities are more accurate and more polarized.