Remaining cycle time prediction: Temporal loss functions and prediction consistency

Abstract

The usefulness of remaining cycle time models for predictive and prescriptive process monitoring depends not only on the overall accuracy of the predictions but also on their earliness: predictions should be as accurate as possible, as early as possible. To give this criterion more weight in model fitting, the paper evaluates three L1 loss functions with temporal decay. All have the property of increasing the weight of residuals from the early stages of a process relative to residuals from later stages but do so to different degrees. The loss functions are used in LSTM networks for training remaining throughout time models of four different business processes based on publicly available event log data sets. Compared to models trained with unweighted L1 loss, the suggested modifications yield small but significant improvements in earliness on out-of-sample data. Neither the unweighted L1 loss nor the modifications led to models with strictly monotonically decreasing predictions of the remaining time.