Local Post-Hoc Explanations for Predictive Process Monitoring in Manufacturing

Explainability is essential for real-world AI adoption in manufacturing: The study demonstrates that simply building accurate predictive models isn't enough—manufacturers need to understand why the model makes specific predictions before they can trust and act on those recommendations in actual production decisions.

Multiple explanation methods provide complementary insights: The research uses two different explanation techniques (Shapley values and ICE plots) to help domain experts examine predictions from different angles, suggesting that relying on a single explanation method may miss important aspects of how an AI system reaches its conclusions.

Bridging the gap between data scientists and domain experts is critical: By combining process mining, machine learning, and explainability methods, the study addresses a key challenge in responsible AI—ensuring that technical insights can be meaningfully communicated to and acted upon by manufacturing professionals who may lack AI expertise.

Local explanations matter for individual decisions: The focus on "local" post-hoc explanations (explaining individual predictions rather than overall model behavior) reflects a practical reality: manufacturing decisions often hinge on understanding why a specific product or process is predicted to fail, not just how the model performs in aggregate.