Learning from the Data to Predict the Process

Abstract Predictive process monitoring (PPM) aims to forecast how a running process instance will unfold in the future, e.g., which activity will be executed next. For this purpose, PPM techniques rely on machine learning models trained on historical event log data. Such models are assumed to learn an implicit representation of the process that accurately reflects the behavior contained in the data, so that they can be used to make correct predictions for new traces with unseen behavior. This capability, called generalization, is fundamental to any machine learning application. However, researchers currently have a limited understanding of what generalization means in a PPM context and how it relates to the characteristics of event logs. In the paper, the authors discuss the generalization capabilities of PPM approaches, focusing on next activity prediction. They develop a framework for generalization in PPM, derived from the understanding of the term in general machine learning. The framework is applied to next activity prediction by developing concrete prediction scenarios, creating corresponding event logs, and using these logs to empirically evaluate the generalization capabilities of state-of-theart models. The evaluation shows that next activity prediction models generalize well in almost all scenarios.