Uncovering Patterns for Local Explanations in Outcome-based Predictive Process Monitoring

Explainable Predictive Process Monitoring aims at deriving explanations of the inner workings of black-box classifiers used to predict the continuation of ongoing process executions. Most existing techniques use data attributes (e.g., the loan amount) to explain the prediction outcomes. However, explanations based on control flow patterns (such as calling the customers first, and then validating the application, or providing early discounts) cannot be provided. This omission may result in many valuable, actionable explanations going undetected. To fill this gap, this paper proposes PABLO (PAttern Based LOcal Explanations), a framework that generates local control-flow aware explanations for a given predictive model. Given a process execution and its outcome prediction, PABLO discovers control-flow patterns from a set of alternative executions, which are used to deliver explanations that support or flip the prediction for the given process execution. Evaluation against real-life event logs shows that PABLO provides high-quality explanations of predictions in terms of fidelity and accurately explains the reasoning behind the predictions of the black box models. A qualitative comparison showcases how the patterns that PABLO derives can influence the prediction outcome, aligned with the early findings from the literature.