Learning semantics-enhanced temporal graph embedding for event prediction
Abstract
Event prediction is an important task for applications such as risk assessment and resource allocation. However, it is non-trivial to model the past due to the complexity and heterogeneity of available data. In recent years, Graph Neural Networks (GNNs) have shown flexibility in processing different forms of data and have been applied for event forecasting because events can be formalized as temporal graph. Although the GNN-based methods have achieved impressive performance, they still have limitations in event prediction task. First, different types of data such as graph and text have not been fully utilized and the contextual information has not been effectively exploited. Secondly, GNNs typically take individual snapshots of temporal graphs as input, which restricts their ability to learn long-term dependencies. To address these problems, we propose a temporal graph embedding learning model, named TGELN. It consists of two major modules, data processor and event predictor. The former processes records of event databases and creates an event-based temporal graph. The latter adopts a novel semantics-enhanced graph neural network that incorporates semantic information as well as other text features to learn temporal graph embedding for event prediction. Extensive experiments are conducted on eight datasets of events between countries, and the results demonstrate the superiority of the proposed model compared to the baseline methods.