The panel discussed where data science needs to go if its results are to be trusted in settings where decisions carry consequences – in industry, administration and critical infrastructures.

Process Mining, Applied to the Inside of a Model

The contribution argued for a change of perspective. Explainability is usually added on top of a finished model: the model stays a black box, and a second method produces a plausible story about its output. Our approach applies process mining – a method established for business processes – to the inside of neural networks, so that the behaviour of a model becomes an object that can be examined, traced and documented rather than merely trusted.

This matters for data science in practice: where a decision has to be justified, a plausible explanation is not enough. What is needed is a faithful account of what the model actually did.

Participants of the ICR’26 conference in Berlin

Conference Contribution

The chair was also represented in the conference programme with a joint paper: Hecht, F. & Grum, M. (2026). Deep Learning-Based Anomaly Detection in Industrial Images: Evaluation and Comparison of Modern Methods, published in the conference proceedings (Springer, Lecture Notes in Networks and Systems).

Read the paper