IGCN: Infected Graph Convolutional Network based Source Identification

Abstract

Source identification has a wide range of applications in daily life, including locating the rumor source in online social networks and finding origins of a rolling blackout in smart grids. Despite great success over the past decade, most prior arts are proposed based an assumption that the underlying propagation model is known in advance. However, this assumption may be impracticable on real scenarios, since it is usually difficult to acquire the actual underlying propagation model. To avoid this limitation, in this paper, we propose the Infected Graph Convolutional Network (IGCN) layer by combining infection network with GCN (Graph Convolutional Network) layers to locate the rumor source without prior knowledge of underlying propagation model. For the first time, we define the problem of source identification as a special graph classification problem with source node as the label. By introducing the feature update method of GCN layer with the idea of attention, we build an IGCN model to adapt the infection networks such that the prediction accuracy on the source is improved under model independent scenarios. We conduct experiments on several real datasets and the results show the superiority of IGCN model to baseline algorlthms.