Predicting Unseen Process Behavior Based on Context Information from Compliance Constraints

Abstract

Predictive process monitoring (PPM) offers multiple benefits for enterprises, e.g., the early planning of resources. The success of PPM-based actions depends on the prediction quality and the explainability of the prediction results. Both, prediction quality and explainability, can be influenced by unseen behavior, i.e., events that have not been observed in the training data so far. Unseen behavior can be caused by, for example, concept drift. Existing approaches are concerned with strategies on how to update the prediction model if unseen behavior occurs. What has not been investigated so far, is the question how unseen behavior itself can be predicted, comparable to approaches from machine learning such as zero-shot learning. Zero-shot learning predicts new classes in case of unavailable training data by exploiting context information. This work follows this idea and proposes an approach to predict unseen process behavior, i.e., unseen event labels, based on process event streams by exploiting compliance constraints as context information. This is reasonable as compliance constraints change frequently and are often the cause for concept drift. The approach employs state transition systems as prediction models in order to explain the effects of predicting unseen behavior. The approach also provides update strategies as the event stream evolves. All algorithms are prototypically implemented and tested on an artificial as well as real-world data set.