Instructions to use ayanjawed/transitpulse-planning-model with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Scikit-learn
How to use ayanjawed/transitpulse-planning-model with Scikit-learn:
from huggingface_hub import hf_hub_download import joblib model = joblib.load( hf_hub_download("ayanjawed/transitpulse-planning-model", "sklearn_model.joblib") ) # only load pickle files from sources you trust # read more about it here https://skops.readthedocs.io/en/stable/persistence.html - Notebooks
- Google Colab
- Kaggle
File size: 4,268 Bytes
abf8393 96ee387 abf8393 0a79302 abf8393 0a79302 abf8393 96ee387 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 | ---
library_name: sklearn
tags:
- tabular-regression
- time-series
- transportation
license: other
---
# TransitPulse planning model
Predicts **hourly bus ridership for one route** (riders per route-hour) from information known in advance:
route, date, hour, schedule and weather. Built for transport planners as a CT-378 Machine Learning course project.
- Live demo (Hugging Face Space, runs the model in the browser): https://huggingface.co/spaces/ayanjawed/TransitPulse
- Live demo (Streamlit): https://ml-project-wuwhhjv3aqsnuxh5x5tdwi.streamlit.app/
- Code, data builder and notebooks: https://github.com/ayanmirza99/ML-Project
## Intended use
- What-if planning for the 24 MTA bus routes it was trained on. It was trained on January to September 2025;
the apps also accept later dates (up to 2027-06-30) with Open-Meteo forecast weather, but those are rough guesses,
especially for March and October to December, which the model never saw.
- **Not** for other routes or other cities (see Limitations).
- The apps show each prediction with a ± 56 riders range (the validation MAE).
## Model
`sklearn.pipeline.Pipeline`: one-hot encoding of `bus_route` and `borough`, then `HistGradientBoostingRegressor`
(learning_rate 0.05, max_iter 500, max_leaf_nodes 63). Selected by validation MAE among models whose saved file is under 50 MB;
a Random Forest was about as accurate but about 110 MB.
Input columns: bus_route, borough, gtfs_n_stops, gtfs_max_stops_per_trip, sched_trips_hr, sched_headway_min, sched_trips_day,
temperature_2m, precipitation, snowfall, wind_speed_10m, hour, day_of_week, is_weekend, hour_sin, hour_cos, is_am_peak, is_pm_peak,
is_holiday, is_school_day.
## Data (three sources plus calendar)
1. **MTA Bus Hourly Ridership: Beginning 2025**, NY Open Data (dataset id `gxb3-akrn`), summed to one row per route per hour.
Licence: the dataset page gives no licence information; it is covered by NY Open Data's terms of use.
2. **MTA GTFS static bus schedules** (mta.info/developers). A 2026 snapshot, not the 2025 service that produced the ridership.
3. **Open-Meteo historical weather** (New York City coordinates, one value per hour for the whole city). Licence: CC BY 4.0. Weather data by Open-Meteo.com.
4. US federal holidays (Python `holidays` package) and a hand-typed, approximate NYC school calendar.
Eight two-week windows between January and September 2025, 24 routes, 64,512 route-hour rows. Missing hours are filled with 0 riders.
Data are from New York because Karachi has no public passenger-count data.
## Evaluation
Time-ordered split by whole days: train 2025-01-13 to 07-14, validation 07-15 to 09-04, test 09-05 to 09-28 (17 days, scored once).
The final model is refit on train plus validation before testing.
| | MAE (riders) | RMSE | R2 |
|---|---|---|---|
| Planning model (this model) | 65.4 | 113.3 | 0.92 |
| Baseline: training average for the route, hour and weekday/weekend | 78.0 | 129.4 | 0.89 |
The average route-hour in the test period has 450 riders, so the error is about 14.5%.
A separate history model (uses recent ridership) reaches MAE 39.0 but needs recent data, so it is not deployed.
Other checks: leave-one-window-out MAE 53.0 vs baseline 68.9; unseen routes MAE 111 vs baseline 132.
## Limitations
- Only 24 routes and eight 2-week windows in one year; the test set is only 17 days, so small differences are not meaningful.
- Some routes changed level during the year (Q58, Q10, Q65 about 35% lower; Q25 about 32% higher from summer). The model does not know about service changes; its error on these routes is about three times larger.
- Weather helps a little, but partly because it hints at the season. It is not evidence that weather causes ridership changes.
- Weather was one city-wide value, and the test used recorded (not forecast) weather, so real use will have higher error.
- Schedule data is a 2026 snapshot; the school calendar is approximate; licence wording was not independently checked.
- New York data only; no claim is made about Karachi.
## How to load
```python
import joblib
model = joblib.load("planning_model.joblib") # needs scikit-learn 1.9.1
```
The input row must contain the columns listed above (see `app/features.py` in the project repository).
|