Research map / Optimization / Optimizers
Contextual Stochastic Optimization for School Desegregation Policymaking
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes redistricting with choices (RWC), a contextual stochastic constraint program that integrates an ML school-choice model (XGBoost/multinomial logit) into school attendance-zone redrawing to minimize SES dissimilarity, applied to Winston-Salem/Forsyth County Schools.
Problem
Redrawing school attendance boundaries to reduce SES segregation is hard because families opt out of assigned schools, and existing redistricting methods ignore how school choice responds to boundary changes.
Approach
Embed a machine-learning school-choice model inside a contextual stochastic optimization (solved via sample average approximation with constraint programming) so that zoning decisions account for predicted student choices, with constraints on travel time, school population change, and contiguity.
Key results
- RWC with XGBoost choice model reduces district SES dissimilarity by 23% (0.596 to 0.459) while rezoning 7,864 students overall, versus 1,870 lower-SES students in R.
- XGBoost choice model achieves 0.742 accuracy and 0.741 weighted F1 on school-choice prediction; top-3 accuracy 0.875.
- Redistricting without choice (R) reduces dissimilarity by 32% (0.596 to 0.403), but ignores opt-outs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use XGBoost instead of the follow-zoned-school baseline | WS/FCS school-choice prediction · WS/FCS school choice, 10-fold cross-validation | accuracy: 0.6533 -> 0.7421 | Helped |
| Use XGBoost instead of the rule-based choice model | WS/FCS school-choice prediction · WS/FCS school choice, 10-fold cross-validation | accuracy: 0.4606 -> 0.7421 | Helped |
| Use XGBoost instead of multinomial logit | WS/FCS school-choice prediction · WS/FCS school choice, 10-fold cross-validation | accuracy: 0.6917 -> 0.7421 | Helped |
| Integrate the XGBoost school-choice model into redistricting | WS/FCS elementary school redistricting · WS/FCS elementary school redistricting | dissimilarity: 0.596 -> 0.459 | Helped |
| Use the RWC choice model instead of the frequency-based choice model | WS/FCS elementary school redistricting · WS/FCS elementary school redistricting | dissimilarity: 0.462 -> 0.459 | Helped |
| Use 50 rather than 30 SAA scenarios | RWC · WS/FCS elementary school redistricting, 100 runs | average dissimilarity: 0.466 -> 0.495 | Hurt |
Surprising findings
- School choice prediction is hard: best macro F1 is only about 0.73 across 41 schools, despite high top-3 accuracy.
- Increasing SAA scenarios from 30 to 50 yielded worse average dissimilarity (0.495 vs 0.466), attributed to optimization difficulty.
Limitations
- Choice model accuracy is modest and many unobserved factors influence choice.
- Only within-district choice is modeled; out-of-district options are ignored.
- Solver does not reach optimality within 12 hours and SAA with more scenarios is less stable.
Experimental setup
Datasets: WS/FCS anonymized geocoded student enrollment data 2019-2023
Benchmarks: WS/FCS student school-choice data (10-fold CV)
Metrics: accuracy top-3 accuracy top-5 accuracy macro F1 weighted F1 dissimilarity index number rezoned opt-out count average driving time
Baselines: Follow zoned school (C^f) Frequency-based rule model (C^rb) Multinomial logit choice model Redistricting without choice (R) Frequency-based redistricting (FR)
Builds on: contextual stochastic optimization sample average approximation XGBoost
Closest papers by meaning
- Achieving Balanced Representation in School Choice with Diversity Goals (AAAI 2025 · similarity 0.81)
- High School Course Scheduling: Student Preferences and Fairness Constraints (Student Abstract) (AAAI 2025 · similarity 0.81)
- Black-Box Optimization with Implicit Constraints for Public Policy (AAAI 2025 · similarity 0.79)
- Allocation Requires Prediction Only if Inequality Is Low (ICML 2024 · similarity 0.79)
- Size-adaptive Hypothesis Testing for Fairness (NeurIPS 2025 · similarity 0.78)
- Implications of Distance over Redistricting Maps: Central and Outlier Maps (AAAI 2024 · similarity 0.78)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.