Causally Reliable Concept Bottleneck Models

Authors: Giovanni De Felice, Arianna Casanova Flores, Francesco De Santis, Silvia Santini, Johannes Schneider, Pietro Barbiero, Alberto Termine

---

We propose Causally reliable Concept Bottleneck Models (C2BMs), a class of concept-based architectures that enforce reasoning through a bottleneck of concepts structured according to a model of the real-world causal mechanisms. We also introduce a pipeline to automatically learn this structure from observational data and unstructured background knowledge (e.g., scientific literature).

Experimental evidence suggest that C2BM are more interpretable, causally reliable, and improve responsiveness to interventions w.r.t. standard opaque and conceptbased models, while maintaining their accuracy.

---

(Tested with Python 3.12 on Ubuntu 22.04.3 LTS)

1. install dependencies with conda:

conda update conda
conda env create -f environment.yml
conda activate c2bm

2. eventually download the datasets: download the required datasets and place them in your CACHE folder ~/.cache/c2bm/{dataset_name}/:

   • asia: is downloaded automatically and saved in ~/.cache/c2bm/asia
   • sachs: is downloaded automatically and saved in ~/.cache/c2bm/sachs
   • alarm: is downloaded automatically and saved in ~/.cache/c2bm/alarm
   • insurance: download .bif file from https://www.bnlearn.com/bnrepository/discrete-medium.html#insurance and save it in ~/.cache/c2bm/bnlearn_bif_datasets/
   • hailfinder: download .bif file from https://www.bnlearn.com/bnrepository/discrete-medium.html#insurance and save it in ~/.cache/c2bm/bnlearn_bif_datasets/
   • celeba: download all required files from https://www.kaggle.com/jessicali9530/celeba-dataset and save them all in ~/.cache/c2bm/CelebA/celeba/
   • penumothorax: dowload the ResNet50 from https://github.com/Soombit-ai/cxr-clip and place the tar in ~/.cache/c2bm/siim_pneumothorax/pretrained_models/ download the test_png train_png folders from https://www.kaggle.com/datasets/abhishek/siim-png-images and place them in ~/.cache/c2bm/siim_pneumothorax/ move the siim_train.csv from this zip to ~/.cache/c2bm/siim_pneumothorax/

3. have a look at the configuration files inside the conf folder. Every configuration file sweep_*.yaml is a sweep configuration used to run experiments on different dataset in sequence. A different sweep is available for each model (e.g., c2bm, cem). A different sweep is available for 2 groups of datasets: synthetic datasets from the bnlearn repository (sweep_{model_name}_bn.yaml) and image datasets (sweep_{model_name}_rw.yaml). The configuration files set the parameters to the ones used in the paper. Comment out the runs you do not want to execute, e.g., less seeds or/and less datasets.

(NOTE: configuring the sweeper using the hydra-list-sweeper package ('grid_params' and 'list_params' notation) prevents from showing possible caught errors during execution. Please use the standard hydra sweeper ('params' notation) for debuggin purposes.)

4. double-check the other settings inside the configuration file, e.g., number of epochs, etc.

5. CAUSAL DISCOVERY:

   • to discover the graph using the automated pipeline:
     • set dataset.load_graph: false.
     • set the override causal_discovery: ges in the configuration defaults.
     • set the override llm: gpt and override rag: standard in the configuration defaults. API-keys are required. Place them in the env.py file.
   • instead, to load the same graph as we discovered in the paper (which we provide for ease of reproducibility):
     • copy the learned_graph/{dataset_name}/graph.pkl file to the respective folder in ~/.cache/c2bm/{dataset_name}/.
     • set dataset.load_true_graph: false.
     • set load_graph: true.
   • instead, to load the true graph, if available:
     • set: dataset.load_true_graph: true.
     • set: dataset.load_graph: false.

6. (optional) to enable logging with wandb, set trainer.logger: wandb and insert you wandb entity in env.py:

7. run the code: python main.py --config-name sweep_{}.yaml

8. Visualization of results: Results will be printed in the terminal. The results are saved in the output folder, organized by the date and time the sweep started. To display the results and the intervention plots:

   • insert the output run paths in the make_plot.py script, following the many (commented out) examples already present in the script.
   • execute python make_plot.py
   • accuracies will be printed in the terminal and the plots will be saved in the plots folder.

Bibtex reference

If you find this code useful please consider citing our paper:

@article{de2025causally,
  title={Causally reliable concept bottleneck models},
  author={De Felice, Giovanni and Flores, Arianna Casanova and De Santis, Francesco and Santini, Silvia and Schneider, Johannes and Barbiero, Pietro and Termine, Alberto},
  journal={arXiv preprint arXiv:2503.04363},
  year={2025}
}