ICDE 2025 | Federated Trajectory Similarity Learning with Privacy-Preserving Clustering

Requirements

• Ubuntu OS
• Python >= 3.5 (Anaconda3 is recommended)
• PyTorch 1.4+
• A Nvidia GPU with cuda 10.2+

Datasets

• We use open-sourced trajectory datasets TDrive, Rome.
• We provided the road network data and map-matching result data for both Tdrive and TaxiRome datasets. For full datasets, please download them from GoogleDrive and put them into ./data/ folder.
• More data can be obtained from OpenStreetMap (OSM)

Reproducibility & Training

1. Data preprocessing (Time embedding and node embedding)

   python preprocess.py

2. Ground truth generating (It will take a while...)

   python spatial_similarity.py
   python temporal_similarity.py

3. Triplets generating

   python data_utils.py

4. Training

   python federated_main.py 

Baseline setting

• FC+FL. The learning rate is set to $0.001$. The model stacks three Fully Connected (FC) layers, each with 64 hidden feature dimensions.
• RNN+FL. The learning rate is set to $0.001$. The number of hidden feature dimensions of the three RNN layers is set to 128.
• GCN+FL. The learning rate is set to $0.001$. The model contains two GCN layers, each with 64 hidden feature dimensions.
• GTS+FL. The hidden feature dimension of POI embedding layer is set to $128$. The parameters $p$ and $q$ in the sampling strategy are set to $1$. In addition, the hidden feature dimension of GNN embedding layer and trajectory embedding layer is selected from the range ${32, 64, 128, 256}$. The learning rate is set to $0.001$.
• ST2Vec+FL. The learning rate is set to $0.001$. The number of hidden feature dimensions of the node2vec layer, GCN embedding layer, date2vec layer, and LSTM layer is set to $64, 64, 64,$ and $128$. The early stop is set to $50$.

The effect of cluster numbers

In the experiment study, we consider 100 clients and find that the proposed F-TSL obtains the best results when setting the number of clusters to 10. To provide a better guideline for how to choose this hyper-parameter, we conduct two additional sets of experiments, setting the numbers of clients to 50 and 80, respectively. The results are shown in above Figures. These results provide evidence that using ten clusters is reasonable. Setting the number of clusters to 10 generally yields better model performance.