Influence Maximization in Social Graphs Based on Community Structure and Node Coverage Gain.
Zhixiao Wang,Chengcheng Sun,Jingke Xi,Xiaocui Li
DOI: https://doi.org/10.1016/j.future.2021.01.025
IF: 7.307
2021-01-01
Future Generation Computer Systems
Abstract:Influence maximization is an optimization problem in the area of social graph analysis, which asks to choose a subset of k individuals to maximize the number of influenced nodes at the end of the diffusion process. As individuals within a community have frequent contact and are more likely to influence each other, community-based influence maximization has attracted considerable attentions. However, this kind of works ignores the role of overlapping nodes in community structure, resulting in performance degradation in seeds selection. In addition, many existing community-based algorithms identify the final seeds only from the selected important communities, or they need to leverage the weights between local spread and global spread of a node. It is difficult to set suitable scales for important communities or to determine the weights for different spread. In this paper, we propose a novel influence maximization approach based on overlapping community structure and node coverage gain. Firstly, social graphs are partitioned into different overlapping communities by the algorithm of node location analysis in topological potential field. Secondly, a node coverage gain sensitive centrality measure is put up to evaluate the influence of each node locally within its belonging communities, which avoids the problem of local spread and global spread. Finally, seed nodes are directly selected by combining the detected community structure with the pre-designed strategy, without important communities identification. The comprehensive experiments under both the Uniform Independent Cascade model and the Weighted Independent Cascade model demonstrate that our proposed approach can achieve competitive influence spread, outperforming state-of-the-art works. Furthermore, our proposed approach exhibits stable performance on graphs with different scales and various structural characteristics.