FFS-Mamba

The PyTorch implementation of my paper Selective Fusion for Self-supervised Trajectory Representation Learning with Dual-View Mamba.

Note

The repository is currently being organized and will be gradually refined after the paper is submitted.

Dataset

For data privacy reasons, we are not permitted to release the original Chengdu and Xi’an datasets. As an alternative, we process Chengdu dataset provided by JGRM for testing, which contains about 200k processed trajectories along with the associated metadata. Experiments on this substitute dataset yield results are similar to those obtained on our internally processed data.

Here is the google drive link for chengdu.

Requirements

pytorch 2.5.1+cu124
mamba-ssm 2.2.6

Pretrain

python ./ffs_pipeline_mlm_contra.py

Downstream Tasks

1 Travel Time Estimation (TTE)

# TTE
python ./ffs_pipeline_tte_crossattn.py 
# or
python ./ffs_pipeline_tte_aug.py

2 Destionation Prediction

Mask 10% tail GPS and corresponding road, use the model to predict the destination road ID.

python ./ffs_pipeline_dp.py

3 Most Simliar Trajectory Search (MSTS)

python ./ffs_pipeline_msts.py