A Spatio-Temporal Graph Neural Network for Process Remaining Time Prediction

Abstract

Predictive process monitoring aims to anticipate the future behavior of ongoing business processes by leveraging historical execution data and real-time event streams. Among its core tasks, remaining time prediction plays a crucial role in enabling proactive decision-making and efficient resource allocation. In this paper, we propose STG-TempNet, a spatio-temporal graph neural network designed to predict the remaining time of process instances. The model integrates three key components: (1) a GINE-based spatial feature extraction module that captures complex control-flow dependencies among activities; (2) a GRU-based short-term dependency module that models local temporal dynamics; and (3) a Transformer-based long-term dependency module that learns global temporal correlations across event sequences. Event logs are transformed into sequences of process instance graphs enriched with structural, temporal, and resource-related attributes. Extensive experiments on four real-world datasets (Sepsis, BPIC2012, BPIC2017, and Traffic Fines) demonstrate that STG-TempNet consistently outperforms state-of-the-art baseline models. It achieves up to a 14.8% improvement over the second-best method and a 32.1% reduction in mean absolute error (MAE) compared with the baseline average, confirming the effectiveness of jointly modeling structural and temporal information for predictive accuracy.