EquiAlgo

Inspiration

EquiAlgo was built to make scholarship allocation more transparent and fair. The baseline model reproduced historical decisions very accurately, but those historical decisions showed a strong regional imbalance. Our goal was therefore not simply to maximize prediction accuracy, but to design a system that remained useful while reducing unfair geographic disparities.

What it does

EquiAlgo evaluates each applicant using two components:

a student score, based on academic performance, relative performance within the study program, and first-generation university status;
a contextual eligibility threshold, adjusted according to family income, weekly work hours, distance from home to campus, and administrative region.

An applicant receives the scholarship when their score is high enough relative to their contextual threshold. Applicants with very low academic performance are filtered out beforehand.

The final model keeps the global scholarship allocation rate within the required 36–44% range while producing much more balanced regional selection rates.

How we built it

We first analyzed the historical dataset and observed that the original model learned from past committee decisions, which themselves contained regional disparities.

We then created an interpretable scoring system:

all applicants start from the same base score;
Cote R can increase or decrease the score;
applicants are ranked by Cote R within their own study program, and this percentile also modifies the score;
first-generation university students receive a small bonus;
contextual factors reduce the minimum threshold required for acceptance.

For the final 4,000 applicants, program percentiles are calculated using the 10,000 historical applicants as the reference population.

We deliberately excluded postal code from the decision process because it was highly redundant with other geographic variables and could act as a strong proxy for region.

Challenges we ran into

The main challenge was avoiding over-correction.

Our first versions strongly favored remote regions because several variables were correlated: region, distance, income, and work hours could all reduce the threshold at the same time.

We therefore iteratively recalibrated the reductions and monitored acceptance rates by region. We also had to balance fairness with the fixed scholarship budget constraint.

Another challenge was making the model interpretable while keeping good predictive performance.

Accomplishments that we're proud of

Our final submission achieved:

89.28% accuracy
88.78% macro F1 score
approximately 42.3% overall scholarship allocation

Regional allocation rates on the final candidate set were also much more balanced:

Bas-Saint-Laurent: 43.65%
Capitale-Nationale: 41.17%
Côte-Nord: 40.53%
Gaspésie–Îles-de-la-Madeleine: 44.71%
Montréal: 42.04%

We are especially proud that the model remains highly interpretable: for every applicant, it is possible to explain which factors affected their score and which factors adjusted their threshold.

What we learned

We learned that removing a sensitive variable is not enough to remove bias. Other variables such as distance, income, work hours, and postal code can act as proxies for geographic location.

We also learned that fairness requires careful calibration: correcting an existing imbalance too aggressively can simply create a new imbalance in the opposite direction.

Finally, we found that an interpretable rule-based model can remain highly competitive in predictive performance while making its decisions easier to audit and explain.

What's next for EquiAlgo

A future version could replace some manually calibrated thresholds with data-driven calibration and sensitivity analysis.

We would also like to contextualize family income more precisely by region, for example by comparing an applicant's income with the income distribution of their own region rather than using the same absolute income brackets everywhere.

Finally, we would evaluate several fairness constraints systematically and construct a Pareto frontier between predictive utility and fairness to select the most defensible operating point.