Adaptive Spatio-Temporal Graph Learning for Bus Station Profiling

This is the implementation of SPA based on PyTorch. The SPA adopts and advances the graph learning structure through a few novel ideas: (1) Designing an adaptive graph learning mechanism to capture the complex and dynamic spatio-temporal dependencies rather than relying on predefined spatio-temporal graphs; (2) Modeling spatio-temporal interactions in shifted spatial graphs to learn fine-grained spatio-temporal features; (3) Employing self-attention mechanism to learn the long-term temporal dependencies preserved in mobility data. We conduct extensive experiments on three real-world spatio-temporal datasets.

structure of the code:

• lib folder: some methods for data loading and processing from AGCRN;
• utils.py: method of loading adjacency graph;
• model.py: implementation of SPA;
• train.py, run.py: train and run the model.

You can use python run.py --dataset PeMSD4 --num_nodes 370 command to run the code.