Exploring Counterfactual Explanations with Temporal and Process Contexts in Predictive Process Monitoring

Andrei Buliga

Predictive Process Monitoring (PPM) is an emerging field within Process Mining that focuses on forecasting the future course of ongoing business cases, enabling organizations to anticipate outcomes and support critical decisions in domains such as finance, healthcare, and public administration. Advances in machine learning have enabled highly accurate predictions, yet these models remain black boxes whose reasoning is inaccessible to end users. This opacity undermines trust and accountability and raises serious challenges for organizations operating under strict regulatory frameworks, such as the GDPR and the EU AI Act, where individuals are entitled to receive understandable explanations of algorithmic decisions that affect them. In the vision of AI-Augmented Business Process Management Systems (ABPMS), process aware explainability is not a peripheral feature but an essential property of next-generation process management tools. Explanations in this context must go beyond generic feature attributions and explicitly respect the temporal and structural nature of business processes. However, most recent efforts to introduce explainability into PPM have transferred techniques from the general Explainable AI domain in a rather naive way. As a result, explanations often lack alignment with process semantics and fail to provide actionable guidance for practitioners. This thesis addresses this gap by focusing on the development of process-aware explana tions for PPM. Among different explanation paradigms, it investigates counterfactual expla nations as a particularly powerful form of actionable reasoning, while also exploring other approaches that extend explanations beyond individual cases. To this end, the thesis makes three main contributions. First, it introduces an evaluation framework tailored to process aware explanations in PPM, defining principled criteria for assessing their quality and useful ness. Second, it develops novel methods for generating explanations that ensure they remain realistic and consistent with the structure and constraints of real processes. Third, it advances the field by demonstrating how explanations can be leveraged in more refined ways: either focusing on summarised process-aware explanations of predictive models , or on the use of explanations for data generation and what-if scenario analysis, enabling the exploration of explanations beyond their initial application. Together, these contributions advance the progress of process-aware explanations as a cornerstone of Explainable Predictive Process Monitoring (XPPM). By moving beyond opaque black-box models and naive explainability techniques, this thesis contributes to the design of next-generation process monitoring systems that are accurate, trustworthy, transparent, and plausible. More broadly, it underscores the importance of combining data-driven learning with process-aware reasoning to ensure the responsible deployment of AI in organizational settings.

Exploring Counterfactual Explanations with Temporal and Process Contexts in Predictive Process Monitoring

(1)