Nostradamus: A novel event propagation prediction approach with spatio-temporal characteristics in non-Euclidean space

Abstract The prediction of event propagation has received extensive attention from the knowledge discovery community for applications such as virus spread analytics, social network analysis, earthquake location prediction, and typhoon tracking. The data describing these phenomena are multidimensional asynchronous event data that affect each other and show complex dynamic patterns in the continuous-time domain. Unlike the discrete characteristics formed by sampling at equal intervals of asynchronous time series, the timestamps of asynchronous events are in the continuous-time field. The study of these dynamic processes and the mining of their potential correlations provide a foundation for applying event propagation prediction, traceability, and causal inference at both the micro and macro levels. Most of the existing methods represent data as being embedded in the Euclidean space. However, when embedding a real-world graph with a tree-likeliness graph, Euclidean space can not visually represent a graph. Inspired by the characteristics of hyperbolic space, we propose a model called Nostradamus to capture the propagation of the events of interest from historical events in a graph via the hyperbolic graph neural Hawkes process with Spatio-temporal characteristics. The Nostradamus’ core concept is to integrate the Hawkes process’s conditional intensity function with a hyperbolic graph convolutional recurrent neural network.