Tableéqui

What it does

Tableéqui audits a scholarship selection model for biases with the ultimate goal being to reduce the biases in the selection of scholarship students and optimise to give an equal opportunity to succeed.

First it determines what are the factors that represent biases. In our case, the main bias to look out for was the location (though it should be relatively easy to apply Tableéqui to other situations). We determined that there was a disproportionnate amount of students from the city receiving scholarships compared to students in remote areas of Quebec (48% of people in the city versus 27% in remote areas).

How we built it

• Leveraged FairLearn alongside custom AUC classifiers to identify and quantify proxy features that leak geographical location.
• Applied Logistic Regression and 500–2000 iteration bootstrap sampling to establish statistically sound confidence intervals for the regional penalty and odds ratios.
• Evaluated multiple fairness interventions including Label Blending (λ), fairlearn's ExponentiatedGradient, and ThresholdOptimizer ## Challenges we ran into
• Optimizing our variables to maximize the number of students that could get in while also not going over budget was difficult, especially as the temptation is to look at the leaderboards and neglet caution for the sake of reckless competition.

Accomplishments that we're proud of

Honestly, just having a working product and making it through the event. It was a great experience, but most of the challenge came down to a lack of time and so producing something decent while using minimal AI tokens (not paying for it) is something we're proud of.

What we learned

I learned a lot about csv file manipulation and pandas. I had previously taken a python course where we briefly looked at all that, but copying code from documentation to make a graph is very different from using the frameworks to extract usefull information from datasets.

What's next for Tableéqui

Hopefully a good night's rest and not too many cavities from all these sugary snacks.

Built With

• pandas
• python
• scikit-learn