Paper: Documenting AI Systems under the EU AI Act: A UML Architectural Framework with Support for Post-Hoc XAI
Link: https://zenodo.org/records/19599421
This repository contains the reference implementations used in the paper.
The purpose of these materials is illustrative and documentary. They demonstrate how heterogeneous AI systems and post-hoc explainability methods can be structured according to the proposed architectural contract, enabling their representation through UML and automated extraction via UMLOOModeler.
For each module presented in the paper, the repository provides:
- Python source code used in the examples
- Expected runtime outputs
- UML class diagrams automatically generated from the source code with UMLOOModeler
These materials allow readers to directly verify the correspondence between:
implementation => UML extraction => compliance-oriented documentation
ClinicalModule Tabular data models using:
- MLP + LIME
- Random Forest + SHAP
ImageModule Image-based classification using:
- CNN + LIME
GeneticModule Sequential genomic data using:
- Bilstm + DeepSHAP
- These examples are not reference architectures and are not intended for benchmarking.
- Model performance and explanation quality are secondary to architectural clarity and traceability.
- The code is intentionally structured to make architectural roles explicit, supporting traceability and auditability.
This repository provides materials demonstrating how the proposed framework can be instantiated in practice.
Details about the UMLOOModeler tool (design goals, usage, and roadmap) are available at:
#1
All materials available in this repository are under the CC BY 4.0 license. You are free to use this material, including for commercial purposes, as long you cite the author.