This repository provides code, data, and analysis scripts for the Alien Zoo framework as used in the study: "Keep Your Friends Close and Your Counterfactuals Closer: Improved Learning From Closest Rather Than Plausible Counterfactual Explanations in an Abstract Setting"
Detailed motivation and rationale are explained in the paper. In short, we provide this framework as a utility to run user studies to perform human level evaluations of counterfactual explanations (CFEs) for machine learning (ML). The presented study investigates whether there is a beneficial effect of adding a plausibility constraint when computing CFEs in the Alien Zoo setting.
Implementation of the experimental framework follows a clear separation into python-based BackEnd, and a javascript based FronEnd.
Very well! Let's go:
There is a list of steps that first need to be done on the BackEnd side before we can go to the Alien Zoo.
- Whatever you do, use Python3.
cd BackEnd
- Install all requirements as listed in
REQUIREMENTS.txt
. - Install
CEML
(Note: required Python 3.6 or higher!):pip install ceml
- Setup an MySQL database (databse name, user name and password of your choice). Then, make sure the credentials and database in
dbmgr.py
(lines 6-10) fit your set up. - Run
python crypt.py
(generates key pair; relevant for encrypting userId information). - Decide which experiment from our study you want to recreate. The default is Experiment 1, but you can change that via the variable
expNo
in line 17 in fileBackEnd/models.py
.
Finally, we can start the server: python server.py
The server is listening on port 8888, so pull up a browser and go to the Alien Zoo under localhost:8888/.
When you're done, stop the running server with Ctrl+C
.
Run BackEnd/db_export.py
to export data from the database. This command generates 6 files from the content of the database. For details on these files and coding of respective data, see here.
Finally, when you are done with the code, run
python BackEnd/reset_database.py user_name user_pw
to reset the database again.
We included the code on how we generated the data as R Markdown documents under BackEnd/modelData: DataSim.Rmd
/ DataSim.pdf
We trained decision tree regression model for each experiment. Lines 23-44 in BackEnd/models.py
show the code used for model training.
Fair enough: The port can be changed in line 19 in file BackEnd/server.py
.
The delays are chosen as used in the reported Experiments. If you want to change them, check lines 84-86 in file FrontEnd/gameUI.js
. Via this file, you can also control other details of the procedure (trials per block, number of blocks, number and placement of attention trials, etc.)
Note that long delays are an effective measure to ensure that participants will really engage with the materials (instead of quickly brushing over everything).
We provide R Markdown documents of the entire statistical evaluation, together with the original user data acquired:
UserData/*.csv
/ uk_plausibleAlienZoo_analysis.Rmd
Note that these files can also be used to recreate all plots (and more!) from the paper.
The UserData
directory also holds .csv-files obtained from the third pilot study, that was used for the a priori power and sample size estimation (see also directory StatisticalEvaluation/PowerAnalysis
).
This work is licensed under a Creative Commons Attribution 4.0 International License.