Some question related to FAT20 paper.

[prateeky2806]

Hi all, I was going through the paper and find it really interesting. I was exploring the DICE code and I was not able to figure out which configurations were used in the FAT20 paper. What would be the appropriate steps for me to reproduce the dpp based method described in the paper (with the corresponding numbers). Any help in this regard would be appreciated.

Thanks, Prateek

[amit-sharma]

The FAccT 2020 paper uses a differentiable model from tensorflow 1.x. The current DiCE library has added support for Tensorflow 2.x and a few other changes for efficiency.

If you are interested in replicating the paper, then I suggest you to install an older version of DiCE (v0.4) and follow the steps in this notebook: https://github.com/interpretml/DiCE/blob/d52094093e1c09563732b6db624fe34d11488b9a/docs/source/notebooks/DiCE_getting_started.ipynb

The notebook also tells you the parameters to use for replication: "The counterfactuals generated above are slightly different from those shown in our paper, where the loss convergence condition was made more conservative for rigorous experimentation. To replicate the results in the paper, add an argument loss_converge_maxiter=2 (the default value is 1) in the exp.generate_counterfactuals() method above. For more info, see generate_counterfactuals() method in dice_ml.dice_interfaces.dice_tensorflow.py."

To install the older version of DiCE, you can simply do, pip install dice-ml==0.4

[prateeky2806]

Thank you very much @amit-sharma for the quick and detailed response, this is really helpful. Maybe another silly question but can you also point me to the part of code which computes the different proposed metrics (validity, proximity, sparsity and diversity)? I am trying to reach a state where I can replicate results in Figure-1 for any of the datasets in the paper.

Thanks, Prateek