Visit frequency prediction

This repository contains the code to train neural network models on visit frequency prediction, i.e., predicting the number of future visits to a newly visited location. Furthermore, we provide the appendix to our paper in this repo.

Unfortunately, our analysis is based on propreitary datasets and can therefore not be reproduced without access to the data.

Installation:

The predict-visits code depends on the graph_trackintel package. The package is available on the MIE Lab GitHub and can be installed with pip in editable mode. To do so:

git clone https://github.com/mie-lab/graph-trackintel.git

In an activated virtual environment, cd into this folder and run

pip install -e .

This will execute the setup.py file and install required dependencies

Preprocessing

First, the graphs are loaded and preprocessed. The data is then saved as a pickle file. One dataset can include several studies. The studies to be included are currently hard-coded in the script. Change the list of studies you want to include in the dataset (variable studies) and execute

python scripts/graph_preprocessing.py -s train_data

This will save a pickle file under data/train_data.pkl with the graphs of all studies that were specified.

Train:

The parameters in the training script are still hardcoded. Change them to the desired model name and the correct dataset names, and run

python scripts/train.py