Evaluation of Explainable Artificial Intelligence for Predictive Process Mining in Education

Process mining leverages event log data to extract valuable knowledge and insights about the underlying processes. Education has embraced process mining, driven by the huge amounts of log data from student activities at the learning management systems (LMS) to enhance processes underlying the event logs of LMS. Educational predictive process mining supports predictions about the future of a running process instance. Predictive efforts are driven by machine learning (ML) and deep learning (DL) approaches. ML and DL approaches are characterized by a high level of efficiency and accuracy in prediction, but also increasing complexity and a low level of explainability. To overcome low explainability, various explainable artificial intelligence (XAI) methods emerged to explain the reasoning process. This study focuses on enhancing explainability in process outcome prediction by examining the properties of interpretability and faithfulness. We evaluate these properties across the primary dimensions of business process data: event attributes, case characteristics, and control flow patterns. Moodle events logs along with various ML and DL algorithms are used to validate the findings. The experiment is conducted to identify which xAI approach is best for educational predictive process mining. This is achieved through the application of key metrics: parsimony, functional complexity, importance ranking correlation, and level of disagreement. These metrics provide a structured approach to evaluating and enhancing the interpretability of predictive models in process mining. Research results in the form of guidelines assist practitioners and researchers in navigating the complex decision-making process by emphasizing the significance of explainability.