Event prediction model combining ordinary differential equation and hypernetworks

Modeling the temporal evolution patterns of network events is crucial for enhancing user experience and creating commercial value. Existing methods combine graph neural networks and time-series models to capture spatial and temporal features. However, they argue that the impact of events on nodes is instantaneous and remains unchanged over time. Furthermore, the utilization of a single fixed model in these methods to handle all events leads to the oversight of the individual characteristics of events. To address these problems, we propose a novel event prediction model combining ordinary differential equations and hypernetworks in this paper called ODEH. First, we leverage graph neural network to construct an ordinary differential equation. The equation effectively captures the continuous information dissemination among nodes and the interaction between node representation channels, facilitating the modeling of the nonlinear and continuous impact of events on nodes. Second, because the instantaneous impact of events on nodes is discrete when they occur, we use a message passing mechanism to effectively model this discrete influence. Finally, we use a hypernetwork to fine-tune the prediction model based on the information of the input events, capturing the individual characteristics of the events and achieving personalized event prediction. Experimental results on four datasets demonstrate ODEH's superiority to baseline methods, highlighting its effectiveness in event prediction. The codes are available at https://github.com/Zhangjunheeee/ODEH.