This repository provides the implementation code for our paper:
ADFormer: Aggregation Differential Transformer for Passenger Demand Forecasting.
The following is a detailed description of each folder.
geo/taxi_zones contains files representing the geographic information of NYC, which can be used for map visualization.
Similarly, xian_hexAddr.txt contains hexagonal grid addresses of Xi'an, which can be mapped to real-world geographic coordinates.
Files such as NYC_Taxi_origin.pkl and NYC_Taxi_destination.pkl are stored as 2D arrays. The "origin/destination" refers to the start and end points of a trip, as the data are obtained from travel orders. The first dimension is the number of spatial units and the second is the number of timesteps. They are used to generate input data in utils/ADFormer_dataset.py.
module.py implements the essential components required by ADFormer.
ADFormer.py forms the core of our method, implementing attention mechanisms across both spatial and temporal dimensions, as well as the ST-Encoder.
ADFormer_config.py is the configuration file used in our experiments, including settings for the dataset, model, and training process.
ADFormer_dataset.py serves several purposes:
(1) It processes raw data (eg. data/xxx_xxx_origin.pkl), such as adding external information and splitting it according to the window and horizon.
(2) It generates the adjacency, distance, and DTW distance matrices based on the geographic and raw data files.
(3) It aggregates spatial units using dtw_mx and produces cluster maps.
ADFormer_trainer.py defines the components and procedures for training, validation, and evaluation.
utils.py includes utility functions used in the experiments.
You can use the following command to train a model:
nohup python main.py --dataset_city XIAN --dataset_name XIAN-Taxi --window 6 --horizon 1 > XIAN-Taxi-6-1.log 2>&1 &
The trained model and results will be saved in the log folder.
Please note: Since reverse transformation may turn zeros into near-zero values (due to normalization using training data only), MAPE can be distorted. Thus, we compute MAPE only for values greater than 5.