Predictive Process Monitoring (PPM) enables organizations to predict future states of ongoing process instances such as the remaining time, the outcome, or the next activity. A process in this context represents a coordinated set of activities that are enacted by a process engine in a specific order. The underlying source of data for PPM are event logs (ex post) or event streams (runtime) emitted for each activity. Although plenty of methods have been proposed to leverage event logs/streams to build prediction models, most works focus on stationary processes, i.e., the methods assume the range of data variability encountered in the event log/stream to remain the same over time. Unfortunately, this is not always the case as deviations from the expected process behaviour might occur quite frequently and updating prediction models becomes inevitable eventually. In this paper we investigate non-stationary processes, i.e., the impact of unseen data variability in event streams on prediction models from a structural and behavioural point of view. Strategies and methods are proposed to incorporate unknown data variability and to update recurrent neural network based models continuously in order to accommodate changing process behaviour. The approach is prototypically implemented and evaluated based on real-world data sets.

Top