Counterfactual Explanations and Model Multiplicity: a Relational Verification View

Repository for the KR 2023 paper "Counterfactual Explanations and Model Multiplicity: a Relational Verification View".

Training scripts

Training and CFX generation using gradient-based algorithms are implemented in the main.py script. To obtain all the parameters, run the following:

python main.py -h

This will return:

positional arguments:
  ds            Dataset name.
  dp            Path to dataset.
  mp            Path where model should be loaded/saved.
  lp            Path where logs should be loaded/saved.
  a             Algorithm used to generate counterfactuals.

optional arguments:
  -h, --help    show this help message and exit
  -train        Controls whether model is trained anew. Default:False.
  -nmods NMODS  Number of models to be considered. Default: 2.
  -nexps NEXPS  Number of cfx to be generated. Default: 1.

Example

Say you want to train two models from scratch, and generate 1 explanation using the wachter method. Type:

python main.py ../datasets/german/ ../models/german/ ../results/german.txt wachter -nexp 1 -train

Models will be trained and saved in "../models/german/". Then, explanations for these models will be generated using the wachter method. A summary of results is printed in "../results/german.txt".

Authors

Francesco Leofante, Elena Botoeva, Vineet Rajani

Do not hesitate to contact us if you have problems using this code, or if you find bugs :)

Citing this work

If you decide to use this code for your experiments, please cite

@inproceedings{LeofanteBotoevaRajani23,
  author       = {Francesco Leofante and Elena Botoeva and Vineet Rajani},
  title        = {Counterfactual Explanations and Model Multiplicity: a Relational Verification View},
  booktitle    = {Proceedings of the 20th International Conference on Principles of Knowledge Representation and Reasoning, {KR} 2023, Rhodes, Greece. September 02 - 08, 2023},
  year         = {2023}
  }