This repository contains the project component for course CSCI-GA 3033-091 (Introduction to Deep Learning Systems) in New York University. The authors are Qingyang Li and Jiahao Chen.

This project aims at discovering the boundaries of business areas in cities and predicting its future variations using trajectory data of taxicabs. More detailed information can be found in the Project Report.

The codes directory contains source code of this project. The whole procedure is compressed into a Jupyter Notebook named pipeline.ipynb. The steps include: data preprocessing; ConvLSTM model construction, training and testing; comparison with other methods; visualization of results.

In addition, convolution_lstm.py includes the definition of the network structure of ConvLSTM, which is the deep learning model we applied to this task. evaluate.py includes python scripts to compute the metrics (Precision, Recall and F1-score) to evaluate the performance of various models.

The orders directory contains trajectory data (plain text) of the original form. The data_paired directory contains trajectory data (plain text) after pre-processing. Please refer to the Project Report for more details.

The count_map directory contains transition cuboids of 20 days, in npy format (numpy matrices). These data serve as the input data of our model. The heat_map directory contains heat maps of the city, also in npy format. These data serve as the label of our model.

The model directory contains saved model parameters. You may use the load_state_dict method provided by PyTorch to load a pretrained model.

You may just follow the steps clearly stated in pipeline.ipynb to run our model. Feel free to replace the orders directory by trajectory data of other cities.