PairNorm: Tackling Oversmoothing in GNNs

Authors: Lingxiao Zhao, Leman Akoglu

ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on real-world graphs demonstrate that PAIRNORM makes deeper GCN, GAT, and SGC models more robust against oversmoothing, and significantly boosts performance for a new problem setting that benefits from deeper GNNs.
Researcher Affiliation Academia Lingxiao Zhao Carnegie Mellon University Pittsburgh, PA 15213, USA {lingxia1}@andrew.cmu.edu Leman Akoglu Carnegie Mellon University Pittsburgh, PA 15213, USA {lakoglu}@andrew.cmu.edu
Pseudocode No The paper describes the PAIRNORM procedure using mathematical equations (Eq. 10 and 11) and an illustration (Figure 2), but does not provide a formally labeled 'Pseudocode' or 'Algorithm' block.
Open Source Code Yes Code is available at https://github.com/Lingxiao Shawn/Pair Norm.
Open Datasets Yes Datasets. We use 4 well-known benchmark datasets in GNN domain: Cora, Citeseer, Pubmed (Sen et al., 2008), and Coauthor CS (Shchur et al., 2018). Their statistics are reported in Appx. A.2.
Dataset Splits Yes For Cora, Citeseer and Pubmed, we use the same dataset splits as Kipf & Welling (2017), where all nodes outside train and validation are used as test set. For Coauthor CS, we randomly split all nodes into train/val/test as 3%/10%/87%, and keep the same split for all experiments.
Hardware Specification Yes We mainly use a single GTX-1080ti GPU, with some SGC experiments ran on an Intel i7-8700k CPU.
Software Dependencies No The paper mentions using specific GNN models and Tensorboard, but does not provide specific version numbers for any software dependencies or libraries (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes Hyperparameters. We choose the hyperparameter s of PAIRNORM from {0.1, 1, 10, 50, 100} over validation set for SGC, while keeping it fixed at s = 1 for both GCN and GAT due to resource limitations. We set the #hidden units of GCN and GAT (#attention heads is set to 1) to 32 and 64 respectively for all datasets. Dropout with rate 0.6 and L2 regularization with penalty 5 10 4 are applied to GCN and GAT. Configurations. For PAIRNORM-enhanced models, we apply PAIRNORM after each graph convolution layer (i.e., after activation if any) in the base model... We run each experiment within 1000 epochs 5 times and report the average performance.