Handling Concept Drift in Predictive Process Monitoring

Predictive process monitoring emerged as a technique to anticipate the outcome of a running instance of a business process. To this end, it first constructs a forecast model based on an encoding of traces of past process executions that are labelled with the prediction target. This model is then used to predict the outcome of a running process instance. However, existing approaches neglect that real-world processes are subject to continuous change, so that prediction models need to adapt to concept drift. In this paper, we take up ideas on incremental learning from general data mining and present a paradigm for predictive process monitoring under concept drift. It is grounded in a systematic experimental study that answers the questions of which encoding of process traces and which incremental learning strategies are particularly suited for predictive monitoring of continuously evolving processes.