This repository contains the official implementation of
Lopardo, G., Precioso, F., & Garreau, D. "Faithful and Robust Local Interpretability for Textual Predictions."
pip install requirements.txt
python -m spacy download en_core_web_lg
To replicate the experiments, simply run:
python3 main.py --dataset DATASET --model MODEL
- DATASET: restaraunts, yelp, tweets, imdb
- MODEL: logistic_classifier, tree_classifier, forest_classifier, distilbert, roberta
The code will then compare the FRED, LIME, SHAP, and Anchors explainers on the given dataset and model, evaluating them on faithfulness, robustness, time, and the proportion of the document used for explainability.
Results will appear in the directory results.
If you just want to apply FRED to explain your model model on a document doc, run
from fred import explainer
explainer = explainer.Fred(class_names=class_names, classifier_fn=model.predict_proba)
exp = explainer.explain_instance(doc)
print(exp.best)
See fred_example.ipynb and fred_saliency.ipynb for counterfactuals and saliency weights tutorials.