Understanding Delay Detection Challenges in Business Processes

Early identification of delays in business processes is crucial for organizations to prevent missed deadlines and service level agreement violations. Predictive process monitoring (PPM) systems, which use historical event logs to forecast remaining case times, have seen advancements with deep learning. However, there's a gap in understanding the inherent challenges of delay detection itself, particularly how models perform across the entire distribution of delays. This research addresses this gap by analyzing the difficulty of delay detection across 14 event logs. The findings indicate that remaining times are typically heavily right-skewed, meaning a small fraction of cases exhibit significantly large delays. While current models accurately predict the most common delays, they perform poorly on these critical, high-delay cases. The study also uncovered pronounced heteroscedasticity, where predictive uncertainty increases proportionally with the magnitude of the delay. Attempts to mitigate the imbalance problem, where high-delay cases are underrepresented, yielded limited benefits. This suggests that the core issue might not be just data imbalance but rather the higher inherent uncertainty associated with predicting large delays. The research concludes that exploiting this correlation through uncertainty-aware modeling could substantially improve the identification of delayed cases, offering a promising direction for future PPM research.