DeepStay: Stay Region Extraction from Location Trajectories using Weak Supervision

The original code to Christian Löwens, Daniela Thyssens, Emma Andersson, Christina Jenkins, and Lars Schmidt-Thieme "DeepStay: Stay Region Extraction from Location Trajectories using Weak Supervision".

This work was shortlisted for the Best Paper/Student Paper Award at ITSC 2023. The preprint version can be found on arXiv.

Requirements

This code was tested on CPU and GPU with the following environment:

• conda:
  • python=3.10.10
  • pytorch=2.0.1
  • numpy=1.24.3
  • pandas=1.5.3
  • tqdm=4.65.0
• conda-forge:
  • geopandas=0.13.0
• pip:
  • tensorboard=2.13.0
  • func_argparse=1.1.1
  • osmnx=1.3.0

Experiment 1: SR Extraction

Our results can be reproduced with these four steps:

1) Datasets

The following commands download and annotate the GeoLife (GL) and ExtraSensory (ES) datasets:

python ./data/gl_preprocessing.py
python ./data/es_preprocessing.py

2) Pre-training

To pre-train DeepStay on the GL dataset, execute the following command:

python ./src/experiment.py --data="gl" --train_data_path="./data/tmp/preprocessed/gl.pkl"

To see all adjustable parameters, call:

python ./src/experiment.py --help

3) Fine-tuning

After pre-training, fine-tune DeepStay on the ES dataset:

python ./src/experiment.py --data="es" --interp_sec=2 --train_data_path="./data/tmp/preprocessed/es_kfold/0_train.pkl" --pretrained_model_path="./trained_models/<PRETRAINED-MODEL-FILENAME>" --no-use_trained_decoder

4) Evaluation & Baselines

Get the final model predictions for the 0_test.pkl test set:

python ./src/evaluation.py get_model_prediction --data="es" --dataset_interp_sec=2 --testdata_filepath="./data/tmp/preprocessed/es_kfold/0_test.pkl" --model_path="./trained_models/<FINETUNED-MODEL-FILENAME>" --output_filepath="<OUTPUT_FILEPATH>"

For kfold evaluation of the baseline algorithms call:

python ./baselines/evaluation.py  --output_filepath="baseline_results.csv"

Experiment 2: Transportation Mode Detection (TMD)

1) Dataset

Download the GL dataset as described in Experiment 1.

2) Training

python ./src/experiment.py --data="tm" --train_data_path="./data/tmp/preprocessed/gl.pkl" --val_frac=0 --test_k=0

3) Evaluation

Get the final model predictions:

python ./src/evaluation.py get_model_prediction --data="tm" --testdata_filepath="./data/tmp/preprocessed/gl.pkl" --model_path="./trained_models/<TMD-MODEL-FILENAME>" --output_filepath="<OUTPUT_FILEPATH>" --tm_test_k=0