Learning to generate synthetic human mobility data: A physics-regularized Gaussian process approach based on multiple kernel learning

Ekin Uğurel, Shuai Huang, Cynthia Chen

Accepted for podium presentation at the International Symposium on Transportation and Traffic Theory (ISTTT) 25

This repo contains object classes, helper functions, and other codes pertinent to the above-referenced paper.

MKL

Physics-regularization

Generating synthetic mobile data

Setup

The code was written in Python 3.10.8.

The following libraries are the minimal required libraries to run the code:

import torch
import gpytorch
import tqdm
import pandas
import numpy
import sklearn
import similaritymeasures
import matplotlib
import pypots
import skmob
import scipy
import geopandas

or you can have everything set up by running:

pip install -r requirements.txt

Usage

For example to train and test the model on the GeoLife dataset execute the following command (insert your own paths after cloning):

python GeoLifeScript.py --input_path_traj {INSERT PATH}/physics-regularized-MTGP/ --input_path_comp {INSERT PATH}/physics-regularized-MTGP/data --output_path {INSERT PATH}/physics-regularized-MTGP/results

Code Structure

• 'geolife' and 'compressed_geolife' folders: contains the raw trajectories of GeoLife users with mode labels and processed trips to show compressed information, respectively.
• GP.py: contains the core exact GP class object used during training
• GeoLifeScript.py: main script for the experiment shown in Section 4.4.
• MKL.py: contains the greedy MKL class object used to learn the composite spatiotemporal kernels
• benchmarkMethods.py: contains an implementation of the LSTM network (and MC simulations of it) used to benchmark our method
• helper_func.py: contains various helper functions referenced in the main script
• methods.py: contains other helpful methods like adding distances, velocities, and bearings onto mobile data
• metrics.py: contains functions to calculate various metrics leveraged in the main script
• models.py: contains the class objects for the full model and its sparse sister
• plots.py: contains various plotting functions
• sgp.py: contains the core class objects for the sparse GP model
• utils.py: contains utility functions referenced in the main script

Licensing

See the LICENSE file for licensing information as it pertains to files in this repository.

Contact

Ekin Uğurel (ugurel [at] uw.edu)