shuaibo919/g-rat

pytorch implementation of the AAAI-24 paper: Shuaibo Hu, Kui Yu. Learning Robust Rationales for Model Explainability: A Guidance-based Approach.

1.Environments

Ubuntu 18.04.6 python 3.9

Requirements

torch==2.0.1 torchmetrics==1.0.1 tqdm==4.65.0 pandas==2.0.3 numpy==1.25.1

2.Datasets

Embedding

run download_embedding.sh in the data/ directory.

BeerAdvocate

run download_data.sh in the data/beer directory.

HotelReview

run download_data.sh in the data/hotel directory.

3.Running Example

1) Real-World Setting:

Run Re-RNP in Appearance aspect(BeerAdvocate) :

sh script/beer/a0.sh

Run FR in Appearance aspect(BeerAdvocate) :

sh script/beer/share/a0.sh

Run G-RAT in Appearance aspect(BeerAdvocate) :

sh script/beer/guide/a0.sh

Similarly, Run Re-RNP in Location aspect(HotelReview) :

bash script/hotel/a0.sh

Other aspects are similar.

2) Synthetic Setting:

First run the corresponding script for saving the skew model(selector/predictor) parameters:

sh script/beer/pretrain_skew_{selector/predictor}_a{0,1,2}.sh

For example, to run the skew-selector experiments on aspect smell, we first run this:

sh script/beer/pretrain_skew_selector_a1.sh

Then, you can run this to get the result of G-RAT on all skew thresholds:

python run_skew_selector.py --model guide --aspect 1

In this repo, we implemented three models that you can choose from using --model:

arg_name	method
sep	Re-RNP
share	FR(Liu et al.,2022)
guide	G-RAT

More details can be founded in run_skew_selector.py and run_skew_predictor.py

3) Parameter Analysis:

The code used in our parameter analysis is all in run_analysis.py, and its usage is similar to that of run_skew_selector.py.

4) Other:

If you want to try different parameters, please modify the script file in the script/ directory.

g-rat's Issues

Hello, have you uploaded your paper to the Internet yet?

I am interested in your work, may I ask if the original paper is now available on the internet?