GFDRR/sirs-public-private

Per-school public-vs-private ownership classifier (SIRS). Benin pilot, open methodology. # sirs-public-private Per-school public-vs-private ownership classifier for the SIRS (School Infrastructure Risk Screening) workstream. Pilot country is Benin; the architecture is designed to extend to Niger, Mali, Guinea, and Ghana with per-country rule files. **Live dashboard:** gfdrr.github.io  ·  **Methodology:** `methodology.md` ## Why this exists The Bank funds public infrastructure. Vulnerability rollups against the "public school stock" need to know which schools are actually public; misclassifying a private school as public inflates the public-stock denominator. Existing school-location datasets (Giga, OpenStreetMap, national EMIS exports) carry public/private signal inconsistently - some have explicit tags, most don't. This project derives a calibrated `ownership_label` per school with an explicit `unknown` band, optimised for high precision on the `private` flag. ## How it works Four-stage pipeline: 1. **Stage A - deterministic labels.** Per-country YAML rules (regex on school names + OSM tag mining) produce a label with an audit trail (`source_signal`, `rule_id`, `confidence`). High precision, modest recall. See `stage-a-rules/README.md`. 2. **Stage B - probabilistic features.** Distance to urban centers, population density (GHSL SMOD), Relative Wealth Index, religious-POI proximity, INFRE department-level priors. See `stage-b-features/`. 3. **Stage C - calibrated classifier.** Logistic model trained on Stage A labels with ADM1-grouped cross-validation. Output is a probability with a three-band discretization (`public`, `private`, `unknown`). See `stage-c-model/`. 4. **Stage D - manual evaluation.** Sampled review against authoritative ministry / EMIS data per country. The full method, model choices, calibration logic, and pilot results are in **`methodology.md`**. ## Repository layout ``` sirs-public-private/ ├── README.md ├── methodology.md # method + Benin pilot results ├── requirements.txt …