EntroMAGNN: An Entropy-Driven Metapath-Based Graph Neural Network for Maritime Emergency Event Prediction

Predicting forthcoming events constitutes a pivotal area within natural language processing and holds significant importance in diverse fields, including emergency management. Current approaches for event predictions involve constructing event graphs and utilizing heterogeneous graph neural networks, specifically meta-path-based graph neural networks. However, these methods necessitate specific prior knowledge to pre-define meta-paths, and current event graphs typically concentrate on basic event elements exclusively, disregarding the context of events, which impedes further improvement in the accuracy of event prediction. Additionally, the scarcity of reliable event datasets for predicting emergencies poses a significant challenge. To address these issues, we propose an Entropy-driven Meta-path-based Graph Neural Network (EntroMAGNN), a novel framework that leverages graph entropy to automatically extract meta-paths from event graphs by assessing the information richness within the graph. We consider the concatenation of structural and semantic representations of each event node in the event graph as its overall representation, conducting event-type prediction through classification and clustering of these representations. To enrich the semantics of events, we augment the events with commonsense knowledge sourced from an external knowledge graph. Lastly, we construct a dataset called the Maritime Emergency Events Dataset (MEED), which is based on authentic maritime emergency reports. Through both supervised and unsupervised experiments on MEED and a publicly available dataset, we demonstrate the effectiveness and feasibility of EntroMAGNN in learning event representation and predicting event types. Compared with other methods, EntroMAGNN achieves an average increase of 0.16 and 0.20 in NMI on the two datasets.