heyyjudes/dissenting-explanations

Dissenting Explanations

This repo accompanies the paper: Dissenting Explanations: Leveraging Disagreement to Reduce Model Overreliance

Data and Preprocessing

We use data from the Folktables and Opinion Spam Dataset. The preprocessing script: data_processing.py can be used to preprocess the opinion spam dataset.

Generating Dissenting Explanations

Global dissenting explanations can be generated using the counter_exp_global.ipynb notebook and local dissenting explanations can be generated using test-time-local-nn.ipynb and test-time-local-svm.ipynb.

Experiment Materials

For our human experiments, we include the questions we used: final_questions.csv.

Citation

@inproceedings{reingold2024dissenting,
title={Dissenting Explanations: Leveraging Disagreement to Reduce Model Overreliance},
author={Reingold, Omer and Shen, Judy Hanwen and Talati, Aditi},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
volume={38},
number={19},
pages={21537--21544},
year={2024}
}