Scholar Prediction

Inspiration

Scholarships can open doors, but historical decisions may reproduce regional disadvantage. We built Scholar Prediction to explore a more transparent allocation process with explicit fairness constraints and a fixed budget.

What it does

Our model allocates 1,600 scholarships among 4,000 candidates using the challenge’s synthetic data. It ranks academic and work contributions while limiting the difference in selection rates between Centre and Remote applicants.

The selected submission achieved 94.73% accuracy and 94.50% macro F1 on the indicative platform. Its Centre/Remote selection-rate gap is 0.1243 percentage points. These platform metrics are separate from the official jury score.

How we built it

We used Python, NumPy, pandas, SciPy, scikit-learn, Matplotlib and Jupyter.

Our final model uses monotone, piecewise-linear academic and work effects. Other variables help fit historical decisions, but their contributions are excluded from ranking. A deterministic allocator enforces the 40% scholarship budget and regional selection constraint.

Challenges we ran into

More complex models, additional derived features and learning-rate experiments did not consistently improve results. We also explored external datasets, but their populations and outcomes did not match the scholarship task closely enough to justify combining them with the training data.

The central challenge was evaluating fairness with potentially biased committee labels. Improving selection parity does not necessarily improve equal opportunity: our selected policy worsened historical-label TPR and FPR gaps, which we disclose in the audit.

Accomplishments that we're proud of

We achieved our best observed indicative result while producing an exactly reproducible prediction file. We respected the scholarship budget, substantially reduced regional selection disparity and documented the model’s assumptions, limitations and trade-offs.

What we learned

More data and more complex models do not automatically produce better predictions. Relevant data, careful validation and transparent assumptions matter.

We also learned that accuracy, selection parity and equal opportunity measure different things. No single score can establish that an allocation process is fair.

What's next for Scholar Prediction

Before real deployment, we would establish an independently reviewed merit rubric and evaluate the model in shadow mode. We would review both awards and refusals across regions, provide explanations and appeals, and monitor outcomes.

Future improvements should be tested against credible independent assessments of merit rather than treating historical decisions as unquestionable ground truth.

Built With

• jupyter
• matplotlib
• numpy
• pandas
• python
• scikit-learn
• scipy