CancelOut GCN Diffusion(CoutGCN): Finding the Influential Spreaders to Diffuse Information

Abstract

Influential spreaders are the most efficient nodes that can communicate with a vast number of users in a network in the easiest and fastest way. Information spreading, viral social marketing, immunization and controlling the spread of rumors, diseases or viruses are some of the top-notch applications for which identifying the source nodes has become a pivotal job. The computational complexities of the state-of-the-art approaches make it challenging to locate these critical nodes. To minimize such loopholes in the existing methods, in this paper we propose a method called CancelOut GCN Diffusion (CoutGCN) to detect the influential spreaders and their effect on diffusion prowess, which is a much simpler process than the previous proposals. Here, we merge the concepts of Graph Convolutional Network (GCN), feature selection, clustering and community detection algorithms to locate the influential nodes to diffuse information throughout the network. The preprocessing layer CancelOut is a feature selection technology, used to determine the node rankings, while GCN is used for node classification. Different clustering algorithms, particularly Kmeans and Kmedoids, and community detection techniques, viz. LGA and Louvain, are implemented to originate the labels for the training of GCN class prediction. Finally we compared all these algorithms with the previous four benchmark methods, such as Adaptive Degreerank, Voterank, K-shell and EnRenew on three real-world social networks to show how and where each method will give the best result according to the application. Experimentally we found that community detection algorithms worked better when fewer initial spreaders were present, as they would be located in more scattered parts of the network. Clustering models outperformed in a denser network when more influential spreaders were chosen.