From Black Box to Glass Box: Evaluating Faithfulness of Process Predictions with GCNNs

Abstract

The volatile digital economy forces enterprises to tap into the potential of data-driven decision-making. Accordingly, proactive management of business processes is increasingly gaining momentum in information system research. In addition to the superior model performance of predictive models, the explainability of deep learning models becomes a crucial requirement for real-world applications. Although recent works on explainable predictive business process monitoring propose various explainability approaches, preliminary research has been conducted on evaluating explanations regarding their faithfulness. Since human-created ground truth for evaluating algorithms is often unavailable or subjective, objective metrics are needed to assess the faithfulness of explanations. We contribute to this research gap by quantitatively and qualitatively investigating the capabilities of different explainability methods for Graph Convolutional Neural Networks in the context of outcome prediction.

Fabian Gwinner

PHD Candidate

My research interests include Machine Learning, especially Deep Learning, XAI, Decision Support Systems