Abstract: Accurate prediction of the remaining time for ongoing business process instances is crucial for making proactive decisions, meeting deadlines, and optimizing resource allocation. This paper introduces a Transformer-based multitask learning framework that improves time prediction by incorporating an auxiliary classification task. The auxiliary task supervises the learning of entity embeddings for categorical attributes in event logs, using cross-entropy loss to guide the model toward more meaningful representations. By combining temporal modeling with supervised embedding learning, the architecture addresses two core challenges in process data: capturing sequence dependencies and understanding categorical relationships. Experiments on a real-world container terminal data set show that the proposed approach reduces the absolute mean error by 48.4% and the square root error by 39.8% compared to established baselines. These results demonstrate that embedding supervision significantly improves predictive performance and can improve the reliability of time-related forecasts in business process monitoring.

Keywords: Business Process Monitoring, Predictive Process Monitoring, Transformer, Entity Embeddings, Multitask Learning, Remaining time prediction

Pages: 2779-2788