changyanchuan/traj-simi-efficiency

[benchmark] Trajectory similarity computation

Trajectory Similarity Measures

Trajectory similarity computation benchmark from an efficiency perspective, including 10 heuristic measures and 6 learned measures. All measures can be run on CPU and GPU.

Code implementation of the paper at VLDB 2024:

@article{chang2023trajectory,
  title={Trajectory Similarity Measurement: An Efficiency Perspective},
  author={Chang, Yanchuan and Tanin, Egemen and Cong, Gao and Jensen, Christian S and Qi, Jianzhong},
  journal={arXiv preprint arXiv:2311.00960},
  year={2023}
}

Requirements

• An Nvidia GPU with CUDA 10.2
• Ubuntu 20.04 LTS with Python 3.7.7
• pip install -r requirements377.txt (See section FAQ below first.)
• Datasets and the snaphots of TrajGAT and RSTS can be downloaded from here, tar -zxvf trajsimi_dataset.tar.gz -C ./data/ and and tar -zxvf trajsimi_snapshot.tar.gz -C ./exp/snapshot
• Compile pyx files cd core_cpu && python setup.py build_ext --inplace --force && cd ..

Quick Start

We provide python scripts that are originally for the empirical studies in our paper, to help get stared. Try the following prompts.

- Compute the similarity values of trajectory pairs without embedding resue:

python test_trajsimi_time.py --dataset porto --effi_exp numtrajs --gpu

python test_trajsimi_time.py --dataset porto --effi_exp numtrajs 

- Compute the similarity values of trajectories in two data sets:

python test_trajsimi_time_DQ.py --dataset porto --effi_exp numtrajs --gpu

python test_trajsimi_time.py --dataset porto --effi_exp numtrajs 

- Query kNN trajectories (learned measures):

python test_knn_time.py --dataset porto --knn_exp dbset_sizes

- Query kNN trajectories (heuristic measures):

See `./query/cpp./readme.md`

- Cluster trajectories:

python test_clustering_time.py --dataset geolife --clustering_exp numtrajs --gpu

python test_clustering_time.py --dataset geolife --clustering_exp numtrajs 

FAQ

Datasets

We have provided the pre-processed Porto, Germany and Geolife datasets (see Requirements Section). Chengdu and Xian datasets are not public datasets, and thus we do not provide them, while they can be downloaded at here under the agreement of the data provider. We also provide the dataset preprocessing scripts in ./preprocessing/. Alternatively, you can pre-process original datasets from the scrath, such as Porto by using ./preprocessing/porto.py.

To use your own datasets, you may need to create your own pre-processing scripts which should be similar to ./preprocessing/porto.py. The meta data of the dataset (see function post_value_updates() in ./config.py) is required as well.

Fail to install packages

If you fail to install the kmedoids==0.4.3 package from pip repositories, see the guide to compile from the source and install.

We specified the cuda version of dgl package in requirements377.py. You may need to modify the version to the one your environment's need or install it separately.

Contact

Email [email protected] if you have any queries.