Concept Driven Retraining Strategies for Predictive Process Monitoring
Student: Alan Alrechah
Title: Concept Driven Retraining Strategies for Predictive Process Monitoring
Supervisor: Dr. Sebastiaan van Zelst
1st Examiner: Prof. Dr. Wil van der Aalst
2nd Examiner: Prof. Dr. Stefan Decker
Predictive process monitoring is a technique that is gaining popularity in the business world. It encompasses the prediction of the outcome, remaining time, or the next event of a business process while it is still in progress, utilizing the data (events) generated by these processes. To this end, a prediction model is learnt based on historical event data from past process executions with the aim of forecasting the parameter of interest (such as outcome, remaining time, or next event) for currently ongoing instances of the process. However, existing approaches typically take the entire history of the process execution as an input, disregarding the dynamic nature of real-world processes that are constantly changing. This highlights the need for prediction models to incorporate these changes in order to provide accurate and reliable predictions. Therefore, in this thesis, we propose a framework that aims to rebalance the training data by detecting change points in the event log and sampling different cases from different sub-logs constructed after applying a concept drift detection method. This method identifies points in time where concept drifts occurred in the original log. The sampled cases are then passed to a prediction model. We evaluate the proposed framework using a synthetic event log and a real-life event log, that both contain different types of concept drifts. We demonstrate that our framework indeed uncovers how different concept drift types influence the accuracy of the prediction.