Trust, but Verify—Post-Hoc Analysis of Industrial Machine Learning via Interpretability Metric Embedding and Surrogate Mapping

Description

Simon Mählkvist, Pontus Netzell, Thomas Helander, and Konstantinos Kyprianidis map the trustworthiness of an industrial model in this 2026 paper. In factory settings, a model can predict well on average yet behave oddly in parts of its input space, under thin data and shifting processes.

Their framework takes interpretability signals, prediction errors and feature credits, squeezes them into a low dimensional map, and fits a smooth stand-in model over it. Behavior becomes a landscape to inspect rather than scattered point explanations. Applied to ceramic heating element production, the map split into a stable region with accurate predictions and a transitional region of ambiguity, matching known process states. The authors offer the map as a base for targeted reliability checks.

Catalog record

Source: https://europepmc.org/articles/PMC13210896