This project contains a Python package gca and Jupyter Notebook to apply Global Causal Analysis (GCA) to the GoEmotions dataset. It corresponds to the paper "Explaining Model Behavior with Global Causal Analysis", accepted at The 1st World Conference on Explainable Artificial Intelligence (xAI 2023).

If you use the gca software, please cite our conference paper:

@inproceedings{Robeer2023,
    author={Robeer, Marcel and Bex, Floris and Feelders, Ad and Prakken, Henry},
    title={{Explaining Model Behavior with Global Causal Analysis}},
    booktitle={Proceedings of the 1st World Conference on Explainable Artificial Intelligence},
    year={2023},
}

Clone this package and install it locally with pip3 install ..

The requirements should be installed automatically. If not, an overview of the requirements is included in requirements.txt.

The experiments are included in GlobalCausalAnalysis.ipynb.

Model finetuning (3 epochs) takes approximately 1.5 hours on a single Tesla T4 GPU (Google Colab). Therefore, in the /data folder we have included a .csv file containing all features for the test split of GoEmotions: $X$ (text), $Z_{task}$, $Z_{robust}$, $Z_{fair}$, $\hat{Y}$ (PRED_label).

In the Notebook, we set the class-wise one-versus-rest contrast, and convert $\hat{Y}$ (PRED_label) into $\hat{Y}{positive}$ (PRED_positive), $\hat{Y}{negative}$ (PRED_negative), $\hat{Y}{ambiguous}$ (PRED_ambiguous) and $\hat{Y}{neutral}$ (PRED_neutral).

FileNotFoundError: [Errno 2] "dot" not found in path.

Install graphviz on your system.

None of PyTorch, TensorFlow >= 2.0, or Flax have been found. Models won't be available and only tokenizers, configuration and file/data utilities can be used.

Install PyTorch, TensorFlow >= 2.0 or Flax to use transformers (only required when training or applying the black-box model).

Cannot set maximum depth of FCI

As of 0.1.3.3, the maximum depth of FCI could not be set. Use pip3 install git+https://github.com/py-why/causal-learn.git@f4c37718a218e8dc8f44fb1ed44a3f4c891a4bab, or a later version of causal-learn if it becomes available in the meantime.