Z-C-for-school / TMN

Under Construction

TMN

Source codes for TMN

Required Packages:

Users need to install the exact trajectory similarity computing package for various trajectory similarity metrics named 'trajectory_distance' which can be downloaded from https://github.com/maikol-solis/trajectory_distance.

Running Procedures:

1. Download data files from the following websites:
   Geolife: https://www.microsoft.com/en-us/download/details.aspx?id=52367
   Porto: https://archive.ics.uci.edu/ml/datasets/Taxi+Service+Trajectory+-+Prediction+Challenge,+ECML+PKDD+2015
2. If you exploit Geolife or Porto dataset, you can use 'preprocess_geolife.py' or 'preprocess_porto.py' to obtain the data.
3. Run 'preprocessing.py' to preprocess the coordinate tuples and generate the ground truth distances between trajectories.
4. To train TMN or Traj2SimVec, run 'subtraj_distance.py' to generate the ground truth distances between sub-trajectories.
5. Run 'train.py' to train the models. Besides TMN, we implement NeuTraj, T3S, SRN and Traj2SimVec. Parameters can be modified in the 'tools/config.py'.
6. To test TMN, you can use 'test_matching_model' function in 'traj_trainer.py' to evalute TMN on trajectory similarity computation task. This function employs the model to generate the trajectory embeddings and conducts the trajectory similarity search task to evalute the performance of TMN using the evalutation metrics mentioned in our paper. By default, 'train.py' will train the model first, and then employ 'test_matching_model' to generate emebddings and conduct search.