Link Prediction Using Temporal Graph Neural Network Model

In this work, we present a Temporal Graph Neural Network (TGNN) architecture specifically designed for link prediction in dynamic graphs. The proposed approach is evaluated on a dynamic social network constructed from internal email communication between employees of Wrocław University of Science and Technology that was collected over a continuous period of 605 days. To capture short-term fluctuations in communication behavior, we introduce the use of very short temporal aggregation windows, down to a single day, for constructing temporal graph snapshots. This fine-grained temporal resolution allows the model to accurately learn evolving interaction patterns and adapt to the dynamic nature of social communication networks. The TGNN model demonstrates consistently high predictive performance, achieving 99.28% ROC-AUC (Receiver Operating Characteristic—Area Under Curve) and 99.17% Average Precision in link prediction tasks. These results confirm that the model is able to distinguish between existing and emerging communication links with high reliability across temporal intervals. The architecture, optimized exclusively for temporal link prediction, effectively utilizes its representational capacity for modeling edge formation processes in time-dependent networks. The findings highlight the potential of focused TGNN architectures and short-time-window modeling in improving predictive accuracy and temporal resolution in link prediction applications involving evolving social or organizational structures.