Unitary Convolutions for Learning on Graphs and Groups

Authors: Bobak Kiani, Lukas Fesser, Melanie Weber

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experimental results here show that unitary/orthogonal variants of graph convolutional networks perform competitively on various graph learning tasks.
Researcher Affiliation Academia John A. Paulson School of Engineering and Applied Sciences, Harvard; e-mail: bkiani@g.harvard.edu John A. Paulson School of Engineering and Applied Sciences, Harvard; e-mail: lukas_fesser@fas.harvard.edu John A. Paulson School of Engineering and Applied Sciences, Harvard; e-mail: mweber@g.harvard.edu
Pseudocode Yes Algorithm 1 Unitary map from Lie algebra... Algorithm 3 Unitary map in Fourier basis
Open Source Code Yes Code available at https://github.com/Weber-GeoML/Unitary_Convolutions
Open Datasets Yes Long Range Graph Benchmark (LRGB) [DRG+22]... Heterophilous Graph Dataset proposed by [PKD+23]... TU Dataset database [MKB+20]... LRGB [DRG+22] Custom See here for license TUDataset [MKB+20] Open Open sourced here Heterophily Data [PKD+23] N/A Data is open source; no license stated in repository
Dataset Splits Yes For a given GNN model, we train on a part of the dataset and evaluate performance on a withheld test set using a train/val/test split of 50/25/25 percent.
Hardware Specification Yes All our experiments were trained on a single GPU (we used either Nvidia Tesla V100 or Nvidia RTX A6000 GPUs).
Software Dependencies No Experiments were run on Pytorch [PGC+17] and specifically the Pytorch Geometric package for training GNNs [FL19]. The paper mentions the software but not specific version numbers.
Experiment Setup Yes Training procedures and hyperparameters are reported in App. G. Reported results in tables are over the mean plus/minus standard deviation.