A New Perspective on "How Graph Neural Networks Go Beyond Weisfeiler-Lehman?"

Authors: Asiri Wijesinghe, Qing Wang

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically verify the strength of our model on different graph learning tasks. It is shown that our model consistently improves the state-of-the-art methods on the benchmark tasks without sacrificing computational simplicity and efficiency.
Researcher Affiliation Academia Asiri Wijesinghe & Qing Wang School of Computing, Australian National University, Canberra, Australia {asiri.wijesinghe, qing.wang}@anu.edu.au
Pseudocode No The paper describes mathematical formulations and equations for its model but does not include structured pseudocode or algorithm blocks.
Open Source Code Yes The implementation can be found at: https://github.com/wokas36/Graph SNN
Open Datasets Yes We use five datasets: three citation network datasets Cora, Citeseer, and Pubmed (Sen et al., 2008) for semi-supervised document classification, one knowledge graph dataset NELL (Carlson et al., 2010) for semi-supervised entity classification, and one OGB dataset ogbn-arxiv from (Hu et al., 2020).
Dataset Splits Yes (1) the standard splits in Kipf & Welling (2017), i.e., 20 nodes from each class for training, 500 nodes for validation and 1000 nodes for testing, for which the results are presented in Table 1; (2) the random splits in Pei et al. (2020), i.e., randomly splitting nodes into 60%, 20% and 20% for training, validation and testing, respectively, for which the results are presented in Table 13 in Appendix B.
Hardware Specification No The paper does not provide specific hardware details (e.g., CPU, GPU models, memory) used for running its experiments.
Software Dependencies No The paper mentions software like "Adam optimizer" and "PyTorch" in the context of experimental setup, but it does not specify version numbers for these or other software dependencies.
Experiment Setup Yes Experimental setup. We use the Adam optimizer (Kingma & Ba, 2015) and λ = 1. For ogbn-arxiv, our models are trained for 500 epochs with the learning rate 0.01, dropout 0.5, hidden units 256, and γ = 0.1. For the other datasets, we use 200 epochs with the learning rate 0.001, and choose the best values for weight decay from {0.001, 0.002, ..., 0.009} and hidden units from {64, 128, 256, 512}. For γ and dropout at each layer, the best value for each model in each dataset is selected from {0.1, 0.2, ..., 0.6}. Graph SNNGAT uses the attention dropout 0.6 and 8 multi-attention heads. Graph SNNGraph SAGE uses the neighborhood sample size 25 with the mean aggregation.