Temporal GNNs for Remaining Time Prediction: An Evaluation

Predicting the remaining time of a business process instance is crucial for enhancing operational decision-making. Sequential and tabular predictive process monitoring (PPM) methods often fail to capture the dynamic, graph-structured nature of real-world processes. In this work, we investigate the effectiveness of temporal graph networks (TGN) for the remaining time prediction. This addresses the critical gap between static graph-based methods and the temporal dynamics in real-world processes. We propose (1) a framework for representing processes as continuoustime temporal graphs, (2) an adapted TGN architecture that leverages both structural and temporal process information to predict remaining times, and (3) a comprehensive empirical evaluation using statistical methods and domain experts comparing our approach against state-of-the-art PPM methods to quantify improvements in remaining time prediction. Our results demonstrate that TGNs outperform state-of-the-art methods, achieving up to $3.7 \%$ improvement in MAE and $13.4 \%$ in RMSE, demonstrating the effectiveness of temporal graph modeling in PPM.