01Market design · Structural modeling
Parcoursup: Stable Choice with Overlapping Constraints
Estimated applicant preferences and compared the current fixed call-order mechanism (PCS) with Stage-Fair College-Proposing Deferred Acceptance (SF-CDA).
- Modeled 886K applicants and 21K programs with a multinomial-logit preference model.
- Benchmarked mechanisms with 1,000 full-market counterfactual simulations.
- Improved scholarship-target compliance by 19.6 percentage points without reducing enrollment or completion speed.
SQLPandasStatsmodelsSimulation
02Data pooling · Economic incentives
Marketplace ML & Online Advertising
Connected predictive accuracy to consumer reach, advertiser value, retailer-level lift, and participation incentives.
- Complementary datasets increased total consumer capture and each retailer’s capture by 50%.
- With asymmetric data ownership, pooling increased the smaller retailer’s audience by 200% while total capture stayed unchanged.
Ridge regressionCounterfactual analysisMarketplace ML
03Machine learning · Certified removal
Efficient Newton Unlearning
Co-developed a scalable second-order method that removes selected data owners’ influence without full retraining.
- Reused a global Hessian across unlearning requests, reducing per-client computation to O(d²).
- Benchmarked non-IID logistic-regression workloads with up to 512 clients and 500 features.
- Achieved lower parameter error than gradient ascent and faster runtime than standard Newton unlearning.
PythonNumPySciPyCertified removal
04Matching theory · Algorithms
Resource-Regional Caps
Designed polynomial-time mechanisms for two-sided matching with shared regional resources and formal guarantees.
- Designed six mechanisms with stability, fairness, efficiency, and incentive guarantees.
- Evaluated 100 simulated markets with 100 students and 10 colleges.
- A specialized mechanism eliminated conflicts when institutions had aligned priorities.
Matching theoryAlgorithm designPython simulation