Where should California put its emergency response centers, and where should responders look in the next hour?
The project uses 1.7M California traffic accidents to answer two questions:
- Placement. Choose k response-center sites that minimize severity-weighted travel time on the real road network.
- Hourly risk. Rank every ~11 km grid cell by how likely an accident is there in the next hour, so a dispatcher with limited attention knows where to look.
An interactive dashboard puts both in front of a planner: live demo. It's on a free host, so the first load can take about a minute to wake up.
Placement: optimizing travel time beats clustering on a map. A severity-weighted p-median on road-network travel times, compared with severity-weighted k-means on latitude and longitude:
| Sites (k) | Method | Weighted mean travel time | Demand reached in 15 min |
|---|---|---|---|
| 10 | k-means | 33.4 min | 23.0% |
| 10 | p-median | 26.1 min | 37.5% |
| 100 | k-means | 11.0 min | 83.9% |
| 100 | p-median | 7.4 min | 84.9% |
Cluster centers look balanced on a map but ignore how roads actually connect. Optimizing travel time directly cut mean response time by 22% at k = 10. Full metrics are in dashboard/outputs/dashboard_erc_metrics.csv.
Hourly risk: ranking the top 4% of cells catches about half of accident locations. The model trains on accidents before 2022 and is tested on 2022 onward, with no look-ahead. A LightGBM ranker adds recency and time-of-week features on top of a smoothed historical-rate baseline:
| Cells checked each hour | Baseline recall | LightGBM recall |
|---|---|---|
| Top 10 | 12.8% | 13.4% |
| Top 50 | 37.1% | 38.2% |
| Top 100 (~4% of the grid) | 51.4% | 53.1% |
The gain over the baseline is small. Where accidents happen is mostly explained by where they happened before, and the model's most important features confirm it: the cell's historical rate, the rate for that hour of the week, and the neighboring cells' rate. The honest takeaway is that a simple, explainable baseline gets most of the way.
US Accidents (2016–2023) by Sobhan Moosavi et al., licensed CC BY-NC-SA 4.0. The raw files (~3 GB) aren't in this repo. Download them from Kaggle and run pipeline/.
Before modeling, the data audit showed what the dataset can't support:
- Coverage grew over time. Records rise more than 4× from 2016 to 2022 because sensor and API coverage expanded, not because accidents did. That rules out year-over-year trend claims.
- Records follow sensors, not crashes. California's 22.5% share of the data reflects where sensors are dense, and rural roads are under-counted.
- Severity labels vary by source. Source1 labels about 90% of records Severity 2, while Source2 labels only about 65% that way. Severity is used as a weight, not as a prediction target.
- Duration is often hard-coded. Many end times are exactly start + 30 min, so duration isn't analyzed.
pipeline/01_clean.py clean the national file, drop high-missing columns
pipeline/02_filter_california.py keep California records
notebooks/01_baselines.ipynb data audit, k-means placement baseline, first risk baseline
notebooks/02_risk_model.ipynb grid, features, LightGBM ranker, Recall@Top-K evaluation
risk_model/ scripts that rebuild the dashboard's risk tables
dashboard/ Dash app (app.py) and the precomputed outputs it reads
results/ model metrics and figures
The p-median solver ran separately on road-network travel times. This repo includes its solutions and metrics (dashboard/outputs/dashboard_erc_*.csv), and the solver code will be added once it's located.
cd dashboard
pip install -r requirements.txt
python app.py # or: gunicorn app:serverThe dashboard reads only the precomputed files in dashboard/outputs/, so it runs without the raw data. render.yaml deploys it to Render as-is.
Team project for UC Berkeley's Analytics Lab course, Spring 2026.