Explainable Artificial Intelligence (VTR UGE 21), day 5 of 5

Practice, Tooling and Governance

Prof. Dr. Utku Kose, Süleyman Demirel University

Governance lab

Part A walks through the risk tiers of the Artificial Intelligence Act of the European Union (EU AI Act) and builds a model card. Part B matches readers with the artefacts they need. Part C computes actionable recourse by hand for a linear credit score.

Part A: Risk tier and model card

This part has two steps. The first walks a system through a simplified version of the risk tiers of the EU AI Act and lists what the provider, the organisation that develops the system, must be able to produce. It is a teaching aid, not legal advice. The second builds a model card from your entries, attaches a fingerprint and lets you download it as a Markdown file. Its entries use terms of the lecture. The test AUC is the area under the receiver operating characteristic (ROC) curve of the model on test data. LIME stands for local interpretable model-agnostic explanations, and TreeSHAP computes Shapley additive explanations (SHAP) for models built from trees. The measured fidelity is the share of the behaviour of the model near a case that the explanation reproduces.

1. Indicative risk tier

Sector of use

Role of the system

Affects natural persons Remote biometric identification Safety component of a regulated product

2. Model card builder

Model name

Test AUC

Local explanation method

Measured fidelity

Intended use

Known limitations, one per line

Continue in Colab, section 3: a model card generated from the model itself.

Part B: Which artefact for which reader?

Each reader needs a different artefact. Match every reader with the artefact that serves them best and press Check the matches.

Reader	Artefact	Feedback

Continue in Colab, section 2: an interactive dashboard for one patient.

Part C: Recourse by hand

A credit model scores applicants with a linear formula and approves a score of zero or more:

score = 2.0 x income (in tens of thousands) - 3.0 x debt ratio + 0.5 x years of credit history - 4.0

An applicant with an income of 18 thousand, a debt ratio of 0.6 and 2 years of history was rejected. Compute three pieces of recourse, each changing one input while the others stay the same.

Continue in Colab, section 5: recourse from a real credit model with a feature that cannot change.