This repository contains the code used in paper "CLEAR: Ranked Multi-Positive Contrastive Representation Learning for Robust Trajectory Similarity Computation"

Requirements

• Ubuntu OS
• Python 3.9.13 (tested)
• PyTorch 1.13.0 (tested)

Preprocessing

We mainly follow t2vec to preprocess the datasets but reproduce all Julia scripts in Python. We suppport two trajectory datasets of different moving objects. They're "porto" and "geolife". Taking "porto" as example, our preprocessing includes several steps:

1. Unify the datasets in diffrent formats:
   python preprocess/preprocess.py -dataset_name "porto"
   Then you'll get a .pkl file called "porto.pkl" in "data/porto".
2. Data augmentation.
   python preprocess/augmentation.py
   Then you'll get a .pkl file named such as "porto_distort_rate_0.2.pkl" in "data/porto/augmentation". Feel free to use multiprocessing :-)
3. Token generation.
   python preprocess/grid_partitioning.py -dataset_name "porto"
   Then you'll get a series of .pkl file named such as "porot_distort_rate_0.2_token.pkl" in "data/porto/token/cell-100_minfreq-50". Again, feel free to use multiprocessing. Meanwhile, the node2vec for the partitioned space will also be done here. You can find the node embedding file "../data/porto/porto_size-256_cellsize-100_minfreq-50_node2vec.pkl"

Training

You can train CLEAR with the following settings.
python main.py -dataset_name "porto" -combination "multi" -loss "pos-rank-out-all" -model_name "clear-DualRNN" -pretrain_mode "pf" -pretrain_method "node2vec" -batch_size 64 -cell_size 100 -minfreq 50 -aug1_name "distort" -aug1_rate 0.4 -aug2_name "downsampling" -aug2_rate 0.4
The trained model will be saved in "{}_checkpoint.pt" and "{}_best.pt". To facilicate the ablation study, they'll be named such as "clear-DualRNN_grid_cell-100_minfreq-50_multi-downsampling-distort-246_pos-rank-out-all_batch-64_pretrain-node2vec-pf_porto_checkpoint.pt" and saved in "data/porto". To reproduce the results of the other variants mentioned in our paper, you can modify the parameters such as combination, loss, batch_size, cell_size and minfreq to corresponding values.

Evaluation

We support three types of evaluation metrics, i.e., "self-similarity", "cross-similarity" and "knn". Taking "self-similarity" as an example, you can follow the next steps to reproduce the results.

1. Prepare experimental dataset.
   python ./experiment/experiment.py -mode data -dataset_name "porto" -exp_list "self-similarity" -partition_method "grid" -cell_size 100 -minfreq 50
   Then you'll get the experimental dataset in "experiment/self-similarity/porto/cellsize-100-minfreq-50".
2. Encode.
   python ./experiment/experiment.py -mode encode -dataset_name "porto" -exp_list "self-similarity" -partition_method "grid" -cell_size 100 -minfreq 50 -combination "multi" -loss "pos-rank-out-all" -batch_size 64 -aug1_name "distort" -aug1_rate 0.4 -aug2_name "downsampling" -aug2_rate 0.4 -model_name "clear-DualRNN" -pretrain_mode "pf" -pretrain_method "node2vec"
   Then you'll get the encoded vector for self-similarity experimental set named with a suffix corresponding to your model, in "experiment/self-similarity/porto/cellsize-100-minfreq-50"
3. Experiment. python experiment.py -mode encode -dataset_name "porto" -exp_list "self-similarity" -spatial_type "grid" -cell_size 100 -minfreq 50 -combination "single" -loss "pos-rank-out-all" -batch_size 64 -aug1_name "distort" -aug1_rate 0.4 -aug2_name "downsampling" -aug2_rate 0.4 -model_name "clear-DualRNN" -pretrain_mode "pf" -pretrain_method "node2vec" Then you'll get the experimental results (.csv file) in "experiment".

We put the unified data file and our trained model of Porto in "data"