Investigation of the Effect of Concept Drift on Data-Aware Remaining Time Prediction of Business Processes

Abstract

Process Mining is a rather new research area in artificial intelligence with handles event logs usually recorded by information systems. Although, remaining time prediction of ongoing business instances has been always a research question in this area, most of the existing literature does not take into account dynamicity of environment and the underlying process commonly known as concept drift. In this paper, a two-phase approach is presented to predict the remaining time of ongoing process instances; in the first phase, future path of process instances is predicted using an annotated transition system with Fuzzy Support Vector Machine probabilities based on case data and in the second phase, the remaining time is predicted by summing up the duration of future activities each estimated by Support Vector Regressor. Finally, a concept drift adaptation method is proposed. To benchmark the proposed prediction method along with the proposed concept drift adaptation method, experiments are conducted using a real-world event log and a simulation event log. The results show that the proposed approach gained 13% improvement on remaining time prediction in case of concept drift.