Haar Graph Pooling

Authors: Yu Guang Wang, Ming Li, Zheng Ma, Guido Montufar, Xiaosheng Zhuang, Yanan Fan

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments in Section 5 demonstrate that the GNN with Haar Pooling achieves state of the art performance on various graph classification and regression tasks.
Researcher Affiliation Academia 1School of Mathematics and Statistics, University of New South Wales, Sydney, Australia 2Max Planck Institute for Mathematics in the Sciences, Leipzig, Germany 3Department of Educational Technology, Zhejiang Normal University, Jinhua, China 4Department of Physics, Princeton University, New Jersey, USA 5Department of Mathematics and Department of Statistics, University of California, Los Angeles 6Department of Mathematics, City University of Hong Kong, Hong Kong.
Pseudocode Yes We attach the algorithmic pseudo-codes for generating the Haar basis on the graph in the supplementary material.
Open Source Code No The paper does not provide a direct link to source code or an explicit statement about its public release. It mentions using PyTorch Geometric, but this refers to a third-party library.
Open Datasets Yes We evaluate Haar Pooling on five widely used benchmark datasets for graph classification (Kersting et al., 2016), including one protein graph dataset PROTEINS (Borgwardt et al., 2005; Dobson & Doig, 2003); two mutagen datasets MUTAG (Debnath et al., 1991; Kriege & Mutzel, 2012) and MUTAGEN (Riesen & Bunke, 2008; Kazius et al., 2005); and two datasets NCI1 and NCI109 (Wale et al., 2008). We also use the QM7 dataset (Blum & Reymond, 2009; Rupp et al., 2012).
Dataset Splits Yes We split the whole dataset into the training, validation, and test sets with percentages 80%, 10%, and 10%, respectively.
Hardware Specification Yes All the experiments use Py Torch Geometric (Fey & Lenssen, 2019) and were run in Google Cloud using 4 Nvidia Telsa T4 with 2560 CUDA cores, compute 7.5, 16GB GDDR6 VRAM.
Software Dependencies No The paper mentions using PyTorch Geometric and the Adam optimizer, but it does not specify exact version numbers for these software dependencies (e.g., PyTorch version, PyTorch Geometric version).
Experiment Setup Yes The hyperparameters of the network are adjusted case by case. We use the Adam optimizer, early stopping criterion, and patience. The early stopping criterion was that the validation loss does not improve for 50 epochs, with a maximum of 150 epochs. The number of nodes in the convolutional layers are all set to 64; the batch size is 60; the learning rate is 0.001 (0.0005 for QM7). The maximal epoch 50 with no early stop for QM7. We repeat all experiments ten times with different random seeds.