Axiomatic Aggregations of Abductive Explanations (Aggrxp)

This repository contains the code and experimental results accompanying the AAAI'24 paper - 'Axiomatic Aggregation of Abductive Explanations'. This work extends XReason to support methods of aggregating abductive explanations for computing feature importance weights.

Getting Started

To install the packages required to execute the code, run

pip install -r requirements.txt

Download datasets and results from data folder and add them to the root folder.

Note that all the datasets are available in datasets/ folder. Folders with names *_ood contain datasets for Lime attack and folders with names *_shapood contain datasets for SHAP attack.

Usage

Aggrxp has a number of parameters, which can be set from the command line. To see the list of options, run (the executable script is located in src):

$ xreason_agg.py -h

Preparing a dataset

All datasets should be added to datasets/ folder. See datasets/compas_ood/ for an example dataset.

Training an XGBoost Model

Before generating abductive explanations, an XGBoost model must be trained on the train dataset.

To train an XGBoost model with T n_estimators, run

python src/xreason.py -t -n T  datasets/traindataset/traindataset.csv

This will generate a temp/someotherpath/*_mod.pkl file.

As an example, you can run the following command to train an XGBoost model with T=100 on compas_ood dataset.

python src/xreason.py -t -n 100 datasets/compas_ood/compas_ood.csv

Generating all abductive explanations

To compute M abductive explanations (AXps) for the predictions of the XGBoost model temp/someotherpath/*_mod.pkl on a test dataset datasets/dataset_name/testdataset.csv, execute the following command.

python src/xreason.py -v -e smt -s z3 --xnum M temp/someotherpath/*_mod.pkl datasets/dataset_name/testdataset.csv

For example, you can generate abductive explanations for the compas_ood_small_test.csv dataset by running the following command.

python src/xreason.py -v -e smt -s z3 --xnum 100 temp/compas_ood/compas_ood_nbestim_100_maxdepth_3_testsplit_0.2.mod.pkl datasets/compas_ood/compas_ood_small_test.csv

We set the value of M sufficiently large to ensure that we get all the possible abductive explanations. The above step will generate a data/somepath1/mwc_expls.pkl file containing all the abductive explanations.

Aggregating abductive explanations

To generate feature importance weights using our methods - Responsibility index, Holler-Packel Index, Deegan-Packel Index, run

python scripts/compute_explanations_single.py data/somepath1/mwc_expls.pkl "dataset_name"

For example, for the compas_ood dataset, the path to explanations file is "data/compas_ood_nbestim_100_maxdepth_3_testsplit_0.2.mod_nbestim_100_maxdepth_3_testsplit_0.2/mwc_expls.pkl" and the command for generating feature importance weights using our methods is

python scripts/compute_explanations_single.py data/compas_ood_nbestim_100_maxdepth_3_testsplit_0.2.mod_nbestim_100_maxdepth_3_testsplit_0.2/mwc_expls.pkl "compas_ood"

Reproducing experimental results of the AAAI'24 paper

All the explanation files generated in our experiments can be found in data/ folder. All the trained models in our experiments can be found in temp/ folder.

For our experiments, we generate datasets for training the OOD classifier for the attack experiments. All the datasets for the LIME attack experiments (compas_ood, compas_ood1, shap_ood, shap_ood1, german_lmodified, german_smodified) and the SHAP attack experiments are stored in datasets/ folder.

To reproduce our results reported in the paper, execute the following steps:

Go to the src directory:

cd src/

Train all the models at once:

source bashscripts/create_models.sh

Given the trained models, run the experimentation script:

source run_experiments.sh

To compute the feature importance weights of biased and corresponding to LIME, SHAP, Responsibility Index, Deegan-Packel Index, and Holler-Packel Index, run

python compute_attack_results.py

The script create_models.sh will train OOD classifiers (XGBoost models) for the compas and german credit datasets. The script run_experiments.sh will invoke the adapted XReason library to generate all AXps, lIME and SHAP explanations on the test datasets and store them in data/ folder. The script compute_attack_results.py will aggregate all the explanations in data/ and compute the feature importance weights of the biased and un-biased features.

Citations

If Aggrxp has been significant to a project that leads to an academic publication, please, acknowledge that fact by citing it:

@misc{biradar2023model,
      title={Model Explanations via the Axiomatic Causal Lens}, 
      author={Gagan Biradar and Vignesh Viswanathan and Yair Zick},
      year={2023},
      eprint={2109.03890},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

elitalobo / aggrxp