Knowing where to go to find customers is the most important question for taxi drivers and ride hailing networks. If demand for taxis can be reliably predicted in real-time, taxi companies can dispatch drivers in a timely manner and drivers can optimize their route decision to maximize their earnings in a given day. Consequently, customers will likely receive more reliable service with shorter wait time. This project aims to use rich trip-level data from the NYC Taxi and Limousine Commission to construct time-series taxi rides data for 63 taxi zones in Manhattan and forecast demand for rides. We will explore deep learning models for time series, including Multilayer Perceptrons, LSTM, Temporal Graph-based Neural Networks, and compare them with a baseline statistical model ARIMAX.
Run the environment.yml file by running the following command on the main repo directory:
conda env create
The installation works for conda==22.9.0. This will install all packages needed to run the data processing code and ARIMAX fitting notebooks with jupyter or Binder.
The model training notebooks were built using Google Colaboratory. The MLP, RNN, and LSTM models are built using pytorch=2.3.1 (i.e., the most updated version of pytorch on Google Colaboratory when we started this project). Therefore, the notebooks training these models should work out-of-the-box if you open them on Colab.
On the other hand, our graph neural networks were built using the torch-geometric-temporal package. This package takes a long time to install and requires some patching due to incompatibility with our version of pytorch. We show how to install a permanent environment in Google Drive in this Colab Notebook. To install the package without a permanent environment, see this Colab Notebook (not recommended).
- 00_a_data_summary.ipynb: Summary of dataset and processing
- 00_b_basic_ts_model.ipynb: Fitting a basic statistical time series model
ARIMAXto test data
- 01_a_final_dataset.ipynb: Notebook generating the dataset we use to train/validate our models (with an 80-20 train-test split)
- 01_b_arimax.ipynb: Notebook training the
ARIMAXmodel on the data
- 02_MLP_for_taxi_dropoff_time_series.ipynb: Notebook training MLP model to the data
- 03_a_rnn_lstm_single_series.ipynb: Notebook training an LSTM model to each taxi zone's time series separately
- 03_b_rnn_lstm_multi_series.ipynb: Notebook training an LSTM model to all the taxi zones simultaneosly
- 03_c_rnn_lstm_multi_series_multivar.ipynb: Notebook training an LSTM model to all the taxi zones simultaneosly and using additional features from the taxi data.
- 03_d_rnn_lstm_validation.ipynb: Contains classes for systematically training and validating baseline, RNN, and LSTM models for final results. Also sets up model that uses month, hours, and day of week embedding layers.
- 04_gnn_fits.ipynb: Notebook training a graphical model on the data
assets: Additional assets unrelated to taxi datadata: Taxi data directorydata_processing: Notebooks for processing the taxi datanotebooks: Notebook files summarizing the data, performing fits, and generating main resultsutils: Custom modules or filesscratch: For unclean files used to develop code
Each directory contains an individual README.md file with more details of directory contents.