Trajectory-User Linking via Multi-Scale Graph Attention Network

This repository provides the implementation of TULMGAT, a novel framework for trajectory-user linking that leverages multi-scale graph attention mechanisms. This work has been accepted by the journal Pattern Recognition (Elsevier).

If you use this code or refer to TULMGAT in your research, please cite our work:

@article{li2025trajectory,
  title={Trajectory-user linking via multi-scale graph attention network},
  author={Li, Yujie and Sun, Tao and Shao, Zezhi and Zhen, Yiqiang and Xu, Yongjun and Wang, Fei},
  journal={Pattern Recognition},
  volume={158},
  pages={110978},
  year={2025},
  publisher={Elsevier}
}

Article link: https://www.sciencedirect.com/science/article/abs/pii/S0031320324007295

• Python 3.10+
• Numpy 1.24.2
• PyTorch 1.13.1
• torch-geometric 2.2.0
• gensim==4.3.2
• scipy==1.12.0

TULMGAT is compared with several traditional trajectory-user linking methods:

1. Prepare dataset: Before running tulmgat.py, ensure that your dataset is properly set up. The dataset should follow the format expected by the data loader.

• PS：We have already provided data and embedding files for Gowalla user=247 in datasets.

2. Data loader: The script datasets/data_loader.py provides functionality to load and preprocess your dataset. Adjust paths as needed.

3. Run TULMGAT: tulmgat.py

1. Scipy Version Compatibility
   Due to maintenance limitations of gensim.models.Word2Vec, only SciPy ≤ 1.12.0 is supported.

2. Grid Splitting Granularity
   The granularity should be adjusted according to the dataset.
   For TULMGAT, the grid size is fixed at 0.01.

3. Data Processing & Leakage Prevention
   The data processing pipeline in data_factory.py avoids data leakage, e.g.:

   • Embeddings are trained only on the training set.
   • Edge relationships are built only from the training set.
   • New grids in the test set can only index <\s> for embedding vectors.

4. POI Processing
   The processing logic for POIs should be consistent with that of grids.
   The graph construction and embedding pipeline can be fully reused.

5. Datasets

   • Gowalla: https://snap.stanford.edu/data/loc-Gowalla.html
   • Brightkite: https://snap.stanford.edu/data/loc-brightkite.html
     Since early works had limited open-source availability, TULMGAT directly uses sequential check-ins extracted from the raw datasets.

6. Performance Notice
   Running TULMGAT directly now will yield better performance than the results reported in the paper.

7. Contact
   For any issues, please contact: liyujie23s@ict.ac.cn