A Novel Dynamic Edge-Adjusted Graph Attention Network for Fire Alarm Data Mining and Prediction

Modern fire alarm systems are essential for public safety, yet they often fail to exploit the wealth of historical alarm data and the complex spatiotemporal dependencies inherent in urban environments. Graph Neural Networks (GNNs) are currently among the most popular methods for handling complex spatiotemporal dependencies. While a range of dynamic GNN approaches have been proposed, many existing GNN-based predictors still rely on a static topology, which limits their ability to fully capture the evolving nature of risk propagation. Furthermore, even among dynamic graph methods, most focus on temporal link prediction or social interaction modeling, with limited exploration in safety-critical applications such as fire alarm prediction. DeaGAT dynamically updates inter-building edge weights through an attention mechanism, enabling the graph structure to evolve in response to shifting risk patterns. A margin-based contrastive learning objective further enhances the quality of node embeddings by distinguishing subtle differences in risk states. In addition, DeaGAT jointly models static building attributes and dynamic alarm sequences, effectively integrating long-term semantic context with short-term temporal dynamics. Extensive experiments on real-world datasets, including comparisons with state-of-the-art baselines and comprehensive ablation studies, demonstrate that DeaGAT achieves superior accuracy and F1-score, validating the effectiveness of dynamic graph updating and contrastive learning in enhancing proactive fire early-warning capabilities.