TGPPN: A Transformer and Graph Neural Network Based Point Process Network Model

The Temporal Point Process (TPP) is applicable in various fields including healthcare, device failure prediction, and social media. It allows for the precise modeling of event occurrences and their associated types, as well as timestamps. Although recent studies have integrated deep learning and reinforcement learning techniques into TPP, most of them focus only on the event sequence without incorporating other fundamental information such as time series information. Moreover, the majority of these studies only predict the information of the next event, which may not be sufficient for practical applications that needs predicting multi-step events. Therefore, we propose the TGPPN model, which employs a Transformer structure to address the multi-step forecasting task, and a graph neural network to handle multi-variable time series in conjunction with event sequences. Our experiments on real-world datasets demonstrate the effectiveness of our proposed model.