Astrologer: Exploiting graph neural Hawkes process for event propagation prediction with spatio-temporal characteristics

Abstract

The prediction of event propagation has received extensive attention from the knowledge discovery community for applications such as social network analysis. The data describing these phenomena are multidimensional asynchronous event data that affect each other and show complex dynamic patterns in the continuous-time domain. The study of these dynamic processes and the mining of their potential correlations provide a foundation for the application of event propagation forecasting.However, conventional forecasting methods often make strong assumptions about the generative processes of the event data that may or may not reflect the reality, and the strong parametric assumptions also restrict the expressive power of the respective processes. Therefore, it is difficult to capture both the temporal and spatial effects of past event sequences.Most of the existing methods capture the intensity function of the Hawkes processes conditioned only on the historical events while ignoring the spatial information and the influences among different events.In this work, we propose the Astrologer, a graph neural Hawkes process that can capture the propagation of events from historical events on graph. The underlying idea of Astrologer is to incorporate the conditional intensity function of the Hawkes processes with a graph convolutional recurrent neural network. Using both synthetic and real-world datasets, we show that, Astrologer can learn the dynamics of event propagation without knowing the actual parametric forms. Astrologer can also learn the dynamics and achieve better predictive performance than other parametric alternatives based on particular prior assumptions.