TG-ODE

This repository provides the official reference implementation of our paper "Temporal Graph ODEs for Irregularly-Sampled Time Series" accepted at the International Joint Conference on Artificial Intelligence (IJCAI) 2024

Requirements

Note: we assume Miniconda/Anaconda is installed, otherwise see this link for correct installation. The proper Python version is installed during the first step of the following procedure.

1. Install the required packages and create the environment

   • conda env create -f env.yml

2. Activate the environment

   • conda activate tgode

How to reproduce our experiments

First, extract the preprocessed data through the command: tar -xvf RESULTS.tar.xz

Then:

export data="" # choose one from ['metrla', 'pems03', 'pems04', 'pems07', 'pems08', 'montevideo', 'heat', 'pow_2_heat', 'pow_5_heat', 'tanh_heat', 'expand_heat', 'reduce_heat', 'gaussian_noise_heat']
export model="" # choose one from ['DCRNN', 'GCRN_LSTM', 'GCRN_GRU', 'TGCN', 'A3TGCN', 'NODE', 'GDE', 'TGODE']
export NUM_CPUS=90 # number of available cpus for the entire experiment
export PERC_GPUS=0.0 # percentage of gpus for one configuration
export CUDA_VISIBLE_DEVICES="" # list of cuda visible devices

• Single-spike heat diffusion

export batch=16
export dir=RESULTS/single_spike/$data/
nohup python3 -u main.py --singlespike --data $data --model $model --batch $batch --savedir $dir --x_scaler StandardScaler >$dir/out_$model_$data 2>$dir/err_$model_$data

• Multi-spike heat diffusion

export batch=16
export dir=RESULTS/multi_spike/$data/
nohup python3 -u main.py --data $data --model $model --batch $batch --savedir $dir --x_scaler StandardScaler >$dir/out_$model_$data 2>$dir/err_$model_$data

• Traffic forecasting

export batch=1
export dir=RESULTS/$data/
nohup python3 -u main.py --data $data --model $model --batch $batch --savedir $dir --x_scaler StandardScaler >$dir/out_$model_$data 2>$dir/err_$model_$data