Supporting Interpretability in Predictive Process Monitoring Using Process Maps

Most predictive process monitoring approaches rely on machine learning techniques. These approaches predict, e.g., the outcome of a process case. As widely known, many machine learning techniques do not inherently provide insights in a useful format for business process experts to interpret the provided predictions and understand the logic used to derive such predictions. Recently, we proposed VisInter4PPM, a business-oriented approach to visually support interpretability in predictive process monitoring. In this paper, we apply VisInter4PPM to a loan request business process, whose behavior is represented in a real-world event log of a financial institution. This is a multiclass prediction problem where requests can be approved, declined, or cancelled. VisInter4PPM relies on the results of the SP-LIME interpreter to generate explanations about the influence of each business process activity on the case outcome. Thus, the SP-LIME results are visually projected onto a BPMN process model. The resulting process map shows which activities contribute to the predicted outcome and to what extent.