Mobility Embedded POIs (ME-POIs)

Official Implementation of Mobility-Embedded POIs: Learning What A Place Is and How It Is Used from Human Movement

Setup

The code was developed with Python 3.13.2. From the repository root, install the dependencies:

python -m pip install -r requirements.txt

Before running either script, update config.json with paths for your data and embeddings. In particular, check data_path, file_name, emb_path, anchor_path, city, and save_dir. The configured paths may need to be changed from the example paths in this repository.

Pretraining

Pretraining loads the visit dataset, text embeddings, and anchor data specified in config.json. Run it from the repository root:

python run_pretrain.py

Set pretraining_strategy to CL or MLM in config.json. When save_pretrained_model is true, the resulting POI embeddings are written to emb_path/poi_embeds.pt.

Evaluation

Evaluation fine-tunes and evaluates POI embeddings for the downstream task in config.json. Run it after pretraining (or after providing a compatible emb_path/poi_embeds.pt):

python run_eval.py

The supported values for downstream_task are open_hours, and is_closed. All three tasks use labels from the processed SafeGraph dataset.

Evaluation also loads the text embedding files for the model names defined in run_eval.py (e5, gtr-t5, gemini, nomic, mpnet, openai-large, and openai-small) from emb_path.