Hierarchical Transformer with Spatio-Temporal Context Aggregation for Next Point-of-Interest Recommendation

This is the Pytorch implementation of STAR-HiT for next Point-of-Interest Recommendation.

Introduction

Spatio-Temporal context AggRegated Hierarchical Transformer (STAR-HiT) is a Transformer-based model for next POI recommendation. STAR-HiT employs stacked hierarchical encoders to recursively encode the spatio-temporal context and explicitly locate subsequences of different granularities, so as to capture the latent hierarchical structure of user movements.

Environmemnt Requirements

• python == 3.6.13
• pytorch == 1.2.0
• numpy == 1.19.2
• pandas == 1.1.5
• sklearn == 0.24.2
• joblib == 1.0.1

Run the Codes

1. Train: replace the data_root in utils/parser.py to where you store the dataset, then run the train.py to train the model. The trained model will be saved in models/.

2. Test: replace the trained_model_path and model_name in predict.py to point to your trained model, then run the predict.py to test the model. The test results will be saved to your trained_model_path.

3. We also provide a model in models/ that is trained on Foursquare NYC. Download the processed datasets below and direct run the predict.py to see the test results.

Datasets

We provide a processed datasets Foursquare NYC. The details are described as follows:

The data_info.csv contains the metadata of the datases, including the number of users, the number of POIs, the max length of check-in sequences.

The {}.pkl files (both of three phases) contains the data for training, validation and test, composed of a set of dictionary as:

{
    "user_id": user id,
    "poi_id_seq_in": check-in poi ids of the input sequence,
    "poi_dist_mat_in": check-in poi distance matrix,
    "poi_timediff_mat_in": check-in time interval matrix (level),
    "poi_out": the target check-in poi id,
    "poi_distance": check-in poi distance, # vector, baselines used only
    "poi_time_interval": check-in time interval (timestamp), # vector, baselines used only
    "poi_timestamp": poi check-in timestamp, # vector, including the target, baselines used only
    "poi_location_quadtree": poi_location_quadtree,  # vector, including the target, baselines used only
    "poi_loc_quad_idx": poi_loc_quad_idx, # including the target, baselines used only
    "poi_dtime_mat_in": time interal matrix (timestamp), # baselines used only
    "poi_seq_out": the target check-in poi sequence # baselines used only
}

Citation

If you find our codes and datasets helpful, please kindly cite the following papers:

@article{STARHiT,
  author    = {Jiayi Xie and
               Zhenzhong Chen},
  title     = {Hierarchical Transformer with Spatio-Temporal Context Aggregation for Next Point-of-Interest Recommendation},
  journal = {ACM Trans. Inf. Syst.},
  volume = {42},
  number = {2},
  year = {2023},
  url = {https://doi.org/10.1145/3597930},
  doi = {10.1145/3597930}}