SPGCL

This repository contains the complete contents of Mining Spatio-Temporal Relations via Self-Paced Graph Contrastive Learning DOI: https://doi.org/10.1145/3534678.3539422, including appendix and poster. I am sorry that the code is not well organized, because I was busy looking for a job during this time, apologize again.

---

This is a Pytorch implementation of SPGCL: Mining Spatio-temporal Relations via Self-paced Graph Contrastive Learning.

Mining Spatio-temporal Relations via Self-paced Graph Contrastive Learning

The Code

Requirements

Following is the suggested way to install the dependencies:

conda install --file SPGCL.yaml

Note that pytorch >=1.10.

Folder Structure

└── code-and-data
    ├── datasets                 # Including datasets
    ├── CONFIG                   # Parameter settings
    ├── lib                      # Contains self-defined modules for our work
    │   |──  adj_from_loc.py     # Build adjacency matrix by KNN
    │   |──  args.py             # Settings about models and loading configure files
    │   |──  dataloader.py       # Dataloaders 
    │   |──  normalization.py    # Data normalizations
    │   |──  utils.py            # Defination of auxiliary functions for running
    │   ├──  layers
    |	│	├──  SPGCL.py    # The core source code of our model SPGCL
    ├── results                  # Saved model and evaluation results are here
    ├── prepare_data.py          # Prepare data before training (run DTW)
    ├── baselines.py             # Some baselines including SVR and ARIMA
    ├── Trainers.py              # Trainer
    ├── train.py                 # Train model
    ├── test.py                  # Test model
    ├── SPGCL.yaml               # The python environment needed for SPGCL
    └── README.md                # This document

Datasets

Step 1: Download PEMS04 and PEMS08 datasets provided by ASTGNN.

Step 2: Before you start running this model, please do:

python prepare_data.py

to prepare DTW matrixs which may cost some time.

Configuration

Step 1: Parameter settings for different datasets are all in ./CONFIG/

Step 2: Important parameters in the configuration are as follows (take PEMS03 as example):

delta = 0.9             # The threshold of positive edges, i.e., \delta^+
delta_negative = 0.2    # The threshold of negative edges, i.e., \delta^-
gamma_1 = 10            # Reweighting hyperparmeter for PU-loss, i.e., \gamma_1
gamma_2 = 20            # Reweighting hyperparmeter for predict loss, i.e., \gamma_2
eta = 0.1               # Prior probablity, i.e., \eta
Q = 0.3                 # Weighting parameters, i.e., \Gamma

Train and Test

Replace PEMS03 with your own dataset name and you can start training and testing your model.

• For train your own model:

  python train.py --data="PEMS03" --data_path="./datasets/PEMS03"
  

• To test the results on the test set:

  python test.py --data="PEMS03" --data_path="./datasets/PEMS03" --conti=True
  

All the parameter settings are in lib.args.py and CONFIG/

Baselines

Our baselines included:

1. Autoregressive Integrated Moving Average model (ARIMA)

2. Support Vector Regression model (SVR)

3. Gated Recurrent Unit model (GRU) (Chung et al.2014)

4. Iterative Deep Graph Learning(IDGL and IDGL-ANCH)(Yu Chen et al.2020), code links IDGL.

5. Spatio-Temporal Graph Convolutional Networks: A Deep Learning Framework for Traffic Forecasting(STGCN) (Wu et al.2020), code links STGCN.

6. Attention Based Spatial-Temporal Graph Convolutional Networks for Traffic Flow Forecasting(ASTGCN and MSTGCN)(Shengnan et al.2019), code links ASTGCN.

7. Land Deformation Predict via Slope-Aware Graph Neural Networks(SA-GNN)(Fan et al.2021)

8. Spatial-Temporal Graph ODE Networks for Traffic Flow Forecasting(STGODE)(Zheng et al.2019), code links STGODE.

9. Adaptive Graph Convolutional Recurrent Network for Traffic Forecasting(AGCRN)(Lei et al.2020), code links AGCRN.

The python implementations of ARIMA/SVR models are in the baselines.py. Code of other baselines (IDGL, IDGL-ANCH, STGCN, ASTGCN, MSTGCN, STGODE, AGCRN) can be found in the corresponding papers and Github links.