<![CDATA[The increasing complexity and interconnectedness of modern security threats, including terrorism, social unrest, and transnational conflicts, pose significant challenges for traditional intelligence and threat detection systems, which struggle to capture both relational and temporal dynamics of evolving security environments. This study aims to develop a predictive framework capable of providing early warnings of emerging security threats by integrating graph-based relational modeling with temporal sequence learning. We propose a hybrid architecture combining Graph Neural Networks (GNN) with bidirectional Long Short-Term Memory (LSTM) networks, enhanced with an attention-based fusion mechanism to jointly model actor interactions and temporal evolution. The framework leverages large-scale event data from GDELT and ACLED spanning 2015–2025, encompassing over 9.8 million events and 14,532 unique actors, and constructs dynamic, attributed security networks to capture multi-dimensional actor relationships. Experimental results demonstrate that the proposed GNN-LSTM model achieves an overall accuracy of 94.3% and an F1-score of 88.3% for critical threat detection, outperforming traditional machine learning baselines and providing early warnings up to nine days in advance. The model also offers interpretability by highlighting influential actors and key relational patterns contributing to threat escalation. These findings suggest that integrating relational and temporal information through hybrid deep learning architectures significantly enhances predictive accuracy and operational utility in security threat assessment, offering a practical tool for proactive decision-making and resource allocation in complex security environments.]]>
Copyrights © 2026
