Learning multivariate Hawkes process via graph neural network

Abstract

We propose the graph recurrent temporal point process (GRTPP), a deep learning model that can estimate simultaneously the probabilities of occurrence for multiple events in continuous time. GRTPP represents the history of multiple events with their event occurring times into a sequence of graphs and embeds the sequence of graphs into node embeddings for each event. Using the node embedding for each event type, GRTPP then estimates the conditional intensity function for the corresponding event by using a neural network. By approximately integrating the estimated intensity functions, GRTPP computes the likelihood of events used to train the model and predicts the next of event. We have verified that GRTPP learns more flexible representation with relational information among events and produces more accurate predictions than state-of-art neural point process models.