Study on the Helpfulness of Explainable Artificial Intelligence

Description

This repository contains the code to generate the questionnaire that was conducted for the sake of our paper Labarta et al.: Study on the Helpfulness of Explainable Artificial Intelligence (2024) as well as the scripts for the analysis of the gathered survey results. New randomly generated questionnaires on the basis of the chosen XAI-methods and dataset (described below) can be easily generated as well.

In our work we specifically examined the question how far the chosen XAI-methods Confidence Scores, LRP, GradCAM, Integrated Gradients, LIME and SHAP enable a user to correctly identify whether a model (AlexNet or VGG16) classified randomly chosen images from the imagenetv2-matched-frequency dataset correctly. For the decision whether to trust or distrust the model, the participants were only given the generated explanation by one of the XAI-methods but not the actual predicted class of the model.
An example question consisting of the original image alongside the explanation image generated by gradCAM can be seen below:

A complete questionnaire that is generated by this repository is represented by a folder containing twelve subfolders representing different questionnaire forms. Each of those subfolders contains the actual questions (original and corresponding explanation images). For further information on the survey design and questionnaire generation procedure we refer to our paper Study on the Helpfulness of Explainable Artificial Intelligence (2022) section III) Methodology D) Survey Design.

Setup

1. Clone the repository.
2. Create a virtual environment, activate it and install the packages defined in requirments.txt via pip install -r requirements.txt.
3. Download the imagenetv2-matched-frequency dataset from http://imagenetv2public.s3-website-us-west-2.amazonaws.com/ and paste the unpacked folder (leave folder name unchanged) in the data folder of the project structure. For more information on this dataset we refer to https://github.com/modestyachts/ImageNetV2.

Usage

Reproduction of questionnaire (forms) used in conducted survey

Remark: All images that are generated by the following steps can be found in the directory "questionnaire_forms_conducted_survey". Thus there is no explicit need for executing the following steps if time and computational resources need to be saved.

1. Run the main.py file to generate the questions for all methods. If questions for certain XAI-method(s) should not be generated, run the main.py file and set the corresponding parameter(s) (--LRP, --gradCam, --LIME, --IntGrad, --CS, --SAHP) to False. For example python main.py --LIME False --CS False will generate all questions except from LIME and Confidence Scores.
2. Find the results in a newly generated folder questionnaire_forms_<currentDate>_<currentTime> in the root directory.
3. TODO: sollen wir noch parameter einfügen der entscheidet, ob bestehende Survey oder neue Survey berechent?

Generation of new randomly generated questionnaire (forms)

1. Run the experiment_creator.py file.
   Explanation: This will generate a new randomly drawn questionnaire plan meeting the conditions defined in our paper. The plan is a 2D-list, where each inner list represents a individual questionnaire form and contains question-tuples of the format (img_idx, model_used, xai_used, is_pred_correct). The 2D-list is saved as a .pickle in data/question_generation.
2. Run the main.py file.
   Explanation: The questionnaire plan generated in step 1 is read and the defined questions are generated iteratively.
3. Find the results in a newly generated folder questionnaire_forms_<currentDate>_<currentTime> in the root directory.

Reproduction of survey analysis results

1. Run the questionnaire_participants_statistics_charts.ipynb notebook to see some visual statistics regarding the participants.
2. Run the questionnaire_metrics_calculator.ipynb notebook to calculate the metrics that are bound to the research questions defined in our paper. Hypotheses testing is done here as well.