Implementation of Wasserstein Globalness, a method to quantify the globalness of an explainer based on the distribution of its explanations over a dataset.
For more details, please see our full paper:
Axiomatic Explainer Globalness via Optimal Transport
Davin Hill*, Josh Bone*, Aria Masoomi, Max Torop, Jennifer Dy
Proceedings of the 28th International Conference on Artificial Intelligence and Statistics (AISTATS) 2025
[Paper]
The notebook Example/cifar10_example.ipynb calculates globalness for a set of precalculated explanations.
- The function wasserstein_globalness() in ./utils/locality_utils.py calculates wasserstein globalness.
- An example implementation of Wasserstein Globalness on CIFAR10 samples is provided in ./Example
Below we detail source code from the manuscript experiments.
Datasets and Black-Box Models: The black-box models evaluated in the experiments section are trained using the code in the Models/blackbox_model_training directory. Datasets are not included in the repository due to file size, however all datasets are publicly available with sources listed in the paper supplement.
- AUC_Experiment contains the code for the experiment in Section 5.1 (incAUC/excAUC/Infidelity comparison)
- ClusterExperiment contains the code for the experiment in Section 5.3 (group experiment)
- JaggedBoundary contains the code for the experiment in Section 5.2 (synthetic dataset)
- time contains the code for estimating computation time for varying number of features (App. D.1)
- Ablation contains the code for the ablation experiment (App. D.1).