Equivariant Subgraph Aggregation Networks

Authors: Beatrice Bevilacqua, Fabrizio Frasca, Derek Lim, Balasubramaniam Srinivasan, Chen Cai, Gopinath Balamurugan, Michael M. Bronstein, Haggai Maron

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

Reproducibility Variable Result LLM Response
Research Type Experimental A comprehensive set of experiments on real and synthetic datasets demonstrates that our framework improves the expressive power and overall performance of popular GNN architectures.
Researcher Affiliation Collaboration Beatrice Bevilacqua Purdue University bbevilac@purdue.edu Fabrizio Frasca Imperial College London & Twitter ffrasca@twitter.com Derek Lim MIT CSAIL dereklim@mit.edu Balasubramaniam Srinivasan Purdue University bsriniv@purdue.edu Chen Cai UCSD CSE c1cai@ucsd.edu Gopinath Balamurugan University of Tuebingen gbm0998@gmail.com Michael M. Bronstein Imperial College London & Twitter mbronstein@twitter.com Haggai Maron NVIDIA Research hmaron@nvidia.com
Pseudocode Yes Algorithm 1 WL Test
Open Source Code Yes Our code is also available.6
Open Datasets Yes We conducted experiments on thirteen graph classification datasets originating from five data repositories: (1) RNI (Abboud et al., 2020) and CSL (Murphy et al., 2019; Dwivedi et al., 2020)... (2) TUD repository (Morris et al., 2020a)... (3) Open Graph Benchmark (Hu et al., 2020) and (4) ZINC12k (Dwivedi et al., 2020).
Dataset Splits Yes we conducted 10-fold cross validation and reported the validation performances at the epoch achieving the highest averaged validation accuracy across all the folds.
Hardware Specification Yes We implemented our approach using the PyG framework (Fey & Lenssen, 2019) and ran the experiments on NVIDIA DGX V100 stations.
Software Dependencies No The paper mentions using 'PyG framework (Fey & Lenssen, 2019)' and 'Weights and Biases framework (Biewald, 2020)' but does not provide specific version numbers for these software dependencies.
Experiment Setup Yes We used Adam optimizer with learning rate decayed by a factor of 0.5 every 50 epochs. The training is stopped after 350 epochs. As for DS-GNN, we implemented Rsubgraphs with summation over node features... We tuned the batch size in {32, 128}, the embedding dimension of the MLPs in {16, 32} and the initial learning rate in {0.01, 0.001}.