W-monster / GRLSTM

[AAAI 2023] GRLSTM: Trajectory Similarity Computation with Graph-based Residual LSTM

This is a PyTorch implementation of GRLSTM.

Author list: Silin Zhou, Jing Li, Hao Wang, Shuo Shang, and Peng Han.

Framework

Experiment

If you want to reproduce the results, please follow the steps below

Attention

1. If you want to use other dataset, please do not apply map matching on trajectories. Only need to do is align the coordinate with road segment or intersection.
2. How to obtain trajectory similarity labels, please refer to Trajectory Similarity Join in Spatial Networks

construct knowledge graph

cd utils
python construct_KG.py

move, rename, create files

1. move file bj(or ny)_e_map.txt into KGE/datasets/beijing(or newyork)/kg and rename e_map.txt

2. move file bj(or ny)_KG_graph.txt into KGE/datasets/beijing(or newyork)/kg and rename train.txt

3. create a empty file and name test.txt

4. create a file named r_map.txt and copy the follow text

road 0
traj_in 1
traj_not_in 2

train knowledge graph embedding

cd KGE
python run_knowledge_representation.py

If you want to change hyper-parameters, please adjust them in KGE/jTransUP/models/base.py.

The output file of KGE is in KGE/log.

construct fusion graph

cd utils
python construct_fusion_graph.py

Train

python Train.py

Validation

python Validation.py

Test

python Test.py