GREAD: Graph Neural Reaction-Diffusion Networks

Authors: Jeongwhan Choi, Seoyoung Hong, Noseong Park, Sung-Bae Cho

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In our experiments with 9 datasets and 28 baselines, our method, called GREAD, outperforms them in a majority of cases.
Researcher Affiliation Academia 1Yonsei University, Seoul, South Korea.
Pseudocode Yes Algorithm 1 How to train our proposed GREAD
Open Source Code Yes Our code is available at https://github. com/jeongwhanchoi/GREAD.
Open Datasets Yes Chameleon, Squirrel (Rozemberczki et al., 2021), iii) Film (Tang et al., 2009), iv, v, vi) Texas, Wisconsin and Cornell from Web KB. We also test on 3 homophilic graphs with high homophily ratios: i) Cora (Mc Callum et al., 2000), ii) Cite Seer (Sen et al., 2008), iii) Pub Med (Yang et al., 2016).
Dataset Splits Yes We report the mean and standard deviation accuracy after running each experiment with 10 fixed train/val/test splits.
Hardware Specification Yes The following software and hardware environments were used for all experiments: ...and i9 CPU, and NVIDIA RTX 3090.
Software Dependencies Yes UBUNTU 18.04 LTS, PYTHON 3.9.12, PYTORCH 1.11.0, PYTORCH GEOMETRIC 2.0.4, TORCHDIFFEQ 0.2.3, NUMPY 1.22.4, SCIPY 1.8.1, MATPLOTLIB 2.2.3, CUDA 11.3, and NVIDIA Driver 465.19
Experiment Setup Yes We fine-tune our model within the hyperparameter search space in Table 12. Our hyperparameter search used the method of W&B Sweeps (Biewald, 2020) with a standard random search with 500 counts. We introduce the best hyperparameter configuration in Tables 13 to 16.