Abstract:
|
A unit circulating in a process is defined by a unique identifier, a sequence of activities and timestamps to record the execution date of activities. This triplet constitutes an individual journey. predictive business process monitoring (PBPM) aims at predicting part or all of the remainder of a unit’s journey. However, time data may come in two ways : 1 timestamp to denote the start of each activity, and in many real processes, 2 timestamps detailing the start and end of each activity. We have found that deep learning methods used for PBPM are confined to single timestamp predictions. We propose a way to predict the beginning and end of each activity in a journey by using a dedicated encoding based on a specific feature engineering to at least encode the data into three features : duration of last activity, time between activities and duration of current activity. Using this feature engineering on Tax et al.’s (2017), we are capable of accurately predicting activities’ beginnings and ends by adding LSTM layers for each time feature and by combining their L1 loss in the final training loss. We benchmarked the performance of this new algorithm using simulated and real process data.
|