Executing the Credit Risk notebook does not generate a de-biased dataset

[biosopher]

Executing the Credit Risk Notebook does not generate a de-biased dataset. The results below are from a brand now GIT pull from the AIF360 repo. As shown at the end, the new "debiased" model now over twice as biased as the original model:

[adrinjalali]

I [almost] concur. I get 0.11 vs 0.21. Is the input data changing? What's the source of these different results?

It may be a good idea to have some of these as tests or doctests in the docstrings (as examples).

[biosopher]

@nrkarthikeyan Can you advise on this? Just sent an email cc'ing the AIOS team. Need this fixed so we can show integration.

[biosopher]

I know the German credit risk dataset is small, but that doesn't explain the odd behavior of AIF360's notebook.  E.g. I looped through a hundred different splits of that dataset looking for one that would de-bias properly. 

In the process I found that for EVERY split of test/train, AIF360 actually generates a more biased dataset than the initial one.  Something else is wrong here.

Here's my notebook showing this bug:  tutorial_credit_scoring_merged.ipynb.zip

[adrinjalali]

The warning shown in the notebook might also be significant. The preprocessing sets the privileged and unprivileged groups, (gender and race I guess), and then in the notebook when the user tries to set the privileged/unprivileged group, it's ignored as a result. That would quite change the results.

[nrkarthikeyan]

Hi all, the optimized pre-processing (used in the original credit scoring tutorial) has a lot of randomness built into it, which will create issues with small datasets. I suggest that we use re-weighing pre-processing to circumvent this issue. Please see the attached notebook:
tutorial_credit_scoring_reweighing.ipynb.zip

The key thing to keep in mind is that reweighing pre-processing works by changing the instance level weights (this is available in dataset.instance_weights). So, the classifier trained on the debiased data should be capable of handling instance level weights. Let me know what you folks think.

[hoffmansc]

Alternatively, we could run it on the Adult dataset which seems to be much more stable and effective.
tutorial_credit_scoring_adult.ipynb.zip

[scottdangelo]

Any update on fixing this? We've a code pattern on developer.ibm.com that uses this notebook, and it shows that using AIF360 makes fairness worse.

[nrkarthikeyan]

We have modified the tutorial and this issue is fixed.