Trajectory–User Linking via Heterogeneous Preference Graph and Dual-Encoder Mutual Distillation (HPG-DEMD)

Citation

If you find our work useful, please cite:

@inproceedings{chen2022MainTUL,
  title={Trajectory–User Linking via Heterogeneous Preference Graph and Dual-Encoder Mutual Distillation},
  author={},
  booktitle={ICDE},
  year={2026}
}

Datasets

Preprocessed datasets are available in data/. See our paper for detailed descriptions.

Dataset	Foursquare_TKY_800	Foursquare_TKY_400	Foursquare_NYC_800	Foursquare_NYC_400	Weeplaces_800	Weeplaces_400
Duration (Days)	320	320	317	317	2,761	2,643
#Categories	239	231	314	304	1,373	1,171
#POIs	39,698	24,526	4,929	4,290	24,649	18,482
#Trajectories	104,413	51,969	33,971	25,419	152,583	75,873
Avg. Length	3.08	3.15	2.92	3.17	2.58	2.62
Density	2.44	1.95	4.28	3.15	6.49	4.37

• (1) Duration (Days): This represents the total time span of the dataset, calculated as the number of days between the first and the last check-in recorded in the entire dataset. For instance, the Foursquare-NYC dataset spans 317 days.

• (2) #Categories: The total number of unique POI categories.

• (3) #POIs: The total number of unique Points of Interest (POIs).

• (4) #Trajectories: The total number of user check-in sequences.

• (5) Avg. Length: It is calculated as the total number of check-ins divided by the total number of trajectories.

• (6) Density: Formally, let $P_v$ be the set of POIs that have been visited by at least one user, and let $U_p$ be the set of distinct users who have visited a specific POI $p$. The density is then defined as: $$ \text{Density} = \frac{1}{|P_v|} \sum_{p \in P_v} |U_p| $$ where $|P_v|$ is the total number of unique visited POIs, and $|U_p|$ is the number of unique users for POI $p$.

Usage

1. Install dependencies

pip install -r requirements.txt

Requirements (requirements.txt):

numpy==2.3.3
pandas==2.3.2
scikit_learn==1.7.2
torch==2.5.1+cu124  # Adjust based on your CUDA version
torch_geometric==2.6.1
tqdm==4.67.1

2. Run the model

cd project
python main.py --dataset foursquare_tky_400

Supported datasets: foursquare_tky_400, foursquare_tky_800, foursquare_nyc_400, foursquare_nyc_800, weeplaces_400, weeplaces_800

Additional parameters can be configured - see main.py for available options.

Note: Adjust torch version in requirements according to your CUDA setup. Our experiments were conducted on RTX3090 GPUs.

Appendix

For additional resources including:

• Computational cost analysis
• Extended experiments

See: appendix.pdf

For code details, please refer to the code comments.