Restaurant Health Grade Predictor (NYC)

Predict whether a restaurant will receive Grade A vs Grade B/C (or worse) on its next NYC health inspection.

Dataset

• NYC OpenData (Socrata) inspections dataset id: 43nn-pn8j
• We fetch the last 3 years via the API and cache a raw snapshot under data/raw/ for reproducibility.

Setup

This repo assumes Python via pyenv (see .python-version).

make setup

Pipeline

One command:

make ml

Step-by-step:

make data
make preprocess
make train
make eval

Results

Latest run (default split: last 20% by inspection_date_t1, threshold=0.5):

• Logistic regression: precision_fail=0.249, recall_fail=0.951, f1_fail=0.394 (n_test=5671)
• Persistence (last observed grade up to t): precision_fail=0.241, recall_fail=0.291, f1_fail=0.263
• RandomForest: precision_fail=0.337, recall_fail=0.393, f1_fail=0.363
• Run make eval to generate reports/confusion_matrix.csv locally

Threshold tradeoff (logistic regression, FAIL class):

• At threshold=0.5: precision_fail≈0.249, recall_fail≈0.947, f1_fail≈0.395
• At threshold=0.7: precision_fail≈0.387, recall_fail≈0.520, f1_fail≈0.444

Raising the threshold improves precision (fewer false alarms) but reduces recall (more missed B/C+).

Threshold tradeoff (RandomForest, FAIL class):

• At threshold=0.5: precision_fail≈0.337, recall_fail≈0.393, f1_fail≈0.363
• At threshold=0.7: precision_fail≈0.356, recall_fail≈0.113, f1_fail≈0.171

For this run, RF’s best F1 among {0.5,0.7} is at 0.5.

Run make eval to generate metrics under reports/ locally (focus is precision/recall/F1 on the “fail” class = B/C+). Reports are generated locally and not tracked in git.

Leakage constraints (summary)

• Rows are inspection events t.
• Features may use information available up to and including t (including grade/score/violations at t).
• Target is grade at t+1.
• Never include any t+1 fields in features; split is time-based on t+1 date.

See docs/leakage.md.

Limitations

• Observational data; policies and inspection practices can change over time.
• Restaurants with sparse history are harder to model.

Ethics note

Predictions should not be used to penalize businesses without due process; outputs are probabilistic and may encode historical biases in enforcement and reporting.

License: MIT.