Locating influence sources in social network by senders and receivers spaces mapping

Influence source locating is important for misinformation detecting and blocking. However, most of existing multiple sources locating methods use only the local structure of the nodes or the shortest path between them. In addition, some methods do not consider the influencing times of the observed nodes and the mutual effect between the influence cascades originated from various sources. These factors hinder these methods from obtaining high-quality multi-source detection results. To overcome such shortcomings, it is necessary to analyze the nodes’ latent structure characteristics in spreading and receiving the influences. This paper presents a representation learning-based approach to detect the influence sources. The algorithm detects the sources using the topological features of the influenced observed nodes. Firstly, a set of candidate sources is constructed by eliminating some nodes which obviously cannot influence the observed ones. The latent spaces of influence senders and receivers are defined to reveal the nodes' features in influence propagation. The nodes are mapped into the mentioned two latent spaces according to their influencing probabilities and influenced times. The latent spaces establish an influence propagation model, where each node's representations can be used to obtain the probability that it becomes a source. To optimize the propagation model, negative sampling method is used to reduce the computation time. Our experimental results on data sets of 5 real networks and 3 synthetic networks demonstrate that precision of the result by our algorithm is on average 10 % higher than those of the other similar algorithms.