Official repository of the paper

Towards symbolic XAI -
Explanation through human understandable logical relationships between features

Installation instruction

When you want to install everything using pip, just write

pip install -r requirements.txt

to install the requirements. And to install this module as a package write

pip install -e .

which will call setup.py, where the -e flag will make sure changes in the repository will also affect the module. You can can call the module afterwards by import symbxai.

Reconstructing the results

Result	Code
Figure 1 a) NLP example	mult-order_query_qualitative_generation.ipynb
Figure 1 b) Vision example	symbXAI_vision_figures.ipynb
Figure 1 b) Quantum chemistry	quantum_chemistry_MDA_Figure1.ipynb
Figure J.19	quantum_chemistry_analyze_MDA_trajectory.ipynb
Figure 2, Figure 3 a)	mult-order_query_qualitative_generation.ipynb
Figure 3 b)	symb_Vis_multi_order_fig2.ipynb
Figure 6	nlp_sst_exp.ipynb
Figure 7	Automatization process on contrastive conjunctions.ipynb
Table 2, Figure G.13, G.14	scripts/perform_perturbation.py and perturbation_analysis.ipynb
Figure 8	-
Figure 9, Table 4	vision_query_search_exp.ipynb
Figure 10, H.16, H.17	mutag_explainations.ipynb
Figure D.11, D.12	nlp_movie_reviews_exp.ipynb
Figure H.15, H.15	NOT CLEAN YET

Other dependencies:

Datasets

Some precomputed data can be found in notebooks/data folder.

For the Facial Expression Recognition task one can download the data under https://www.kaggle.com/datasets/msambare/fer2013

Citation

If you want to cite us please use this BibTex entry

@article{schnake2025symbxai,
    title = {Towards symbolic XAI — explanation through human understandable logical relationships between features},
    journal = {Information Fusion},
    volume = {118},
    pages = {102923},
    year = {2025},
    issn = {1566-2535},
    doi = {https://doi.org/10.1016/j.inffus.2024.102923},
    author = {Thomas Schnake and Farnoush {Rezaei Jafari} and Jonas Lederer and Ping Xiong and Shinichi Nakajima and Stefan Gugler and Grégoire Montavon and Klaus-Robert Müller},
    keywords = {Explainable AI, Concept relevance, Higher-order explanation, Transformers, Graph neural networks, Symbolic AI},
}