Interval-Based Remaining Time Prediction for Business Processes

Uncertainty is an unavoidable factor in predictive business process monitoring, especially in terms of remaining time prediction. However, existing methods only give a precise time as the result, which fails to consider and reveal the uncertainty of ongoing processes. As a novel attempt to add quantified uncertainty into process monitoring, this paper proposes a model that provides comprehensive predictive information. Specifically, an interval-based time predictor is constructed to make both an optimistic and a pessimistic forecast of the remaining time for business processes. In addition, a clustering-based method is used to extract trace patterns as prior knowledge to optimize interval prediction. We investigate LSTM networks as an approach to construct qualifying time intervals as well as different trace embedding and clustering methods. Our model achieves acceptable results on real-life event logs according to the measurement of coverage-width criterion.