Scrutinizing XAI with linear ground-truth data

This repository contains all the experiments presented in the corresponding paper: "Scrutinizing XAI using linear ground-truth data with suppressor variables".

We use Pipfiles to create Python environments. Since we use innvestigate to create the saliency maps, and this framework uses particular dependencies, there is one extra Pipfile included in the saliency_method folder.

In three steps we can reproduce the results: (i) we generate the ground truth data, (ii) train the linear models and apply the XAI methods, (iii) run the evaluation steps and generate plots.

Generate data

Set the parameter pattern_type=0 to use the signal pattern and suppressor combination analyzed in the paper (see image above). Use pattern_type=3 to generate the data, used to produce the result in the supplementary material.

python -m data.main --path data/config.json 

Run the experiments of model agnostic XAI methods

Update the data_path parameter of the agnostic_methods/conf.json with the path to the freshly generated pickle file containing the ground truth data.

python -m agnostic_methods.main_global_explanations --path agnostic_methods/config.json

Run experiment for sample based explanation, which will take a couple hours, depending on your machine. Here update the data_path of the file agnostic_methods/config_sample_based.json.

python -m agnostic_methods.main_sample_based_explanations --path agnostic_methods/config_sample_based.json

Run experiment of saliency methods

Create a new Python environment, and run the experiments for heat-mapping methods by running through the notebook, change the file_path variable in the notebook.

compute_explanations_heatmapping.ipynb

Run evaluation and generate plots

Update the parameter data_path and results_paths of the config.json. Add the data path and the paths to the artifacts of the experiments.

python run_evaluation_and_visualization.py --path config.json