Motif-Aware Riemannian Graph Neural Network with Generative-Contrastive Learning

Authors: Li Sun, Zhenhao Huang, Zixi Wang, Feiyang Wang, Hao Peng, Philip S. Yu

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

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical results show the superiority of Mofit RGC.
Researcher Affiliation Academia 1North China Electric Power University, Beijing 102206, China 2Beijing University of Posts and Telecommunications, Beijing 100876, China 3Beihang University, Beijing 100191, China 4Department of Computer Science, University of Illinois at Chicago, IL 60607, USA ccesunli@ncepu.edu.cn; penghao@buaa.edu.cn; psyu@uic.edu
Pseudocode Yes Algorithm 1: Optimizing Motif RGC
Open Source Code Yes Codes are given in https://github.com/Riemann Graph/Motif RGC.
Open Datasets Yes We choose 4 public datasets: Cora, Citeseer and Pubmed (Yang, Cohen, and Salakhutdinov 2016), and Airport (Chami et al. 2019).
Dataset Splits No The paper mentions using well-known public datasets but does not explicitly provide details about the specific train/validation/test splits, such as percentages or sample counts for reproduction, beyond implying standard usage of these datasets.
Hardware Specification No The paper does not explicitly specify any hardware details such as GPU models, CPU types, or memory used for the experiments. It only mentions general aspects like 'Py Torch'.
Software Dependencies No The paper mentions 'Py Torch' and optimizers 'Riemannian Adam (B ecigneul and Ganea 2019)' and 'Adam (Kingma and Ba 2015)' but does not provide specific version numbers for these or any other software components, which is required for reproducibility.
Experiment Setup Yes In our model, the convolution layer is stacked twice, and MLP has 2 hidden layers. The number of learnable factors is 3 with the curvatures κ1 = 1, κ2 = 1 and κ3 = 1 as default. α = 2 for contrastive learning. To initialize Riemannian features, we first initialize X RN dm, where N and dm are number of nodes and factor dimension. Then, we have Xm = X 2 κm X max Gdm κm in the factor manifold, where X max is the maximum norm of the rows.