Predicting information diffusion using the inter- and intra-path of influence transitivity

Predicting information diffusion helps grasp the overall preference of user interactions, facilitating applications such as public opinion analysis and online marketing. Existing approaches sample the information cascade network into several independent paths or subgraphs to learn cascade representations, resulting in information loss regarding the social influence of nodes and dynamics between cascades across different temporal stages. To address such problems, we design a deep learning-based model (named I3T) using Inter-and Intra-Path of Influence Transitivity to predict the incremental popularity of information diffusion in information networks. First, we leverage a graph neural network (GNN) to aggregate the node information of the local neighbors. Then, we sample the information cascade into a group of sequences using DeepWalk and update the node embedding with GNN and DeepWalk simultaneously, which embodies both the inter-path and intra-path of influence transitivity. Next, we exploit bi-directional long short-term memory (Bi-LSTM) to extract structural features and apply gated recurrent unit (GRU) to extract temporal features. Finally, we learn the structural factor weight under the temporal guidance of the attention mechanism. The results of comprehensive experiments conducted on two representative datasets demonstrate the preeminence of I3T over existing state-of-the-art approaches.