Online Model Generation for Scalable Predictive Process Monitoring

Abstract

Predictive process monitoring techniques are intended to forecast outcomes of running process instances. This is achieved through the use of predictive models inferred from past event logs. However, the use of such procedures in scenarios where there is an initial lack of previous data or concept drift can be especially challenging. To overcome those limitations, this paper is focused on enabling online model generation, addressing problems such as uncertainty about process completeness. Further, a scalable streaming system based on Apache Flink platform is built and applied on an incident management system dataset in order to assess prediction performance. The results presented show the capacity of these techniques to support predictive process monitoring.