Learning Bi-directional Social Influence in Information Cascades using Graph Sequence Attention Networks

The online information cascades have spatial and temporal characteristics. Retweets in cascades have bi-directional social influence and temporal delay, e.g., a hub node influenced by a root node will further increase the exposure of information, enhance the influence of the root node, and cause subsequent retweets after a certain time. Existing deep learning approaches mostly consider only the decay effects of social influence, ignoring the bi-directional dependency and delay effects between graphs at different moments. Therefore, a novel method is presented here, namely the Graph Sequence Attention Networks (GSAN), which addresses the bi-directional attention mechanism to learn temporal dynamics of social influence, as well as the cascading structure. A graph transformer block is designed to learn the complicated dependencies of spatial and temporal features in cascades. The proposed method could achieve state-of-the-art performance and large improvements in popularity prediction compared to strong baselines in real-world datasets.