Social Influence Prediction by a Community-based Convolutional Neural Network

Abstract

Learning social influence between users on social networks has been extensively studied in a decade. Many models were proposed to model the microscopic diffusion process or to directly predict the final diffusion results. However, most of them need expensive Monte Carlo simulations to estimate diffusion results and some of them just predict the size of the spread via regression techniques, where people who will adopt the information becomes unknown. In this work, we regard the prediction of final influence diffusion results in a social network as a classification problem to avoid expensive simulations with knowing the final adopters. We first address the problem on a deep neural network and utilize the diffusion traces to train the network. Furthermore, we propose a community-based convolutional neural network to capture the information of local structure with the aforementioned network. The proposed model is referred to as the Social Influence Learning on Community-based Convolutional Neural Network, SIL-CCNN. In the experiment, SIL-CCNN shows the promising results in both synthetic and real-world datasets.