Pseudo-Riemannian Graph Convolutional Networks
Authors: Bo Xiong, Shichao Zhu, Nico Potyka, Shirui Pan, Chuan Zhou, Steffen Staab
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the effectiveness of Q-GCN on graph reconstruction, node classification and link prediction. |
| Researcher Affiliation | Academia | Bo Xiong University of Stuttgart Stuttgart, Germany Shichao Zhu IIE, Chinese Academy of Sciences Beijing, China Nico Potyka Imperial College London London, United Kingdom Shirui Pan Griffith University Queensland, Australia Chuan Zhou AMSS, Chinese Academy of Sciences Beijing, China Steffen Staab University of Stuttgart University of Southampton Stuttgart, Germany |
| Pseudocode | No | The paper describes methods mathematically and verbally but does not include a clearly labeled pseudocode or algorithm block. |
| Open Source Code | Yes | Source code is open available at https://github.com/xiongbo010/QGCN. |
| Open Datasets | Yes | We benchmark graph reconstruction on four real-world graphs including 1) Web-Edu [34]: a web network consisting of the .edu domain; 2) Power [35]: a power grid distribution network with backbone structure; 3) Bio-Worm [36]: a worms gene network; 4) Facebook [37]: a dense social network from Facebook. We consider four benchmark datasets: Airport, Pubmed, Citeseer and Cora, where Airport is airline networks, Pubmed, Citeseer and Cora are three citation networks. |
| Dataset Splits | Yes | For node classification, we use the same dataset split as [39] for citation datasets, where 20 nodes per class are used for training, and 500 nodes are used for validation and 1000 nodes are used for testing. For Airport, we split the dataset into 70/15/15. For link prediction, the edges are split into 85/5/10 percent for training, validation and testing for all datasets. |
| Hardware Specification | No | The paper states 'The time complexity is the same as a vanilla GCN given by O(|V |dd0+|E|d0), where |V | and |E| are the number of nodes and edges, d and d0 are the dimension of input and hidden features. The computation can be parallelized across all nodes. Similar to other non-Euclidean GCNs [32, 33, 17], the mapping from manifolds to the tangent space consume additional computation resources, compared with Euclidean GCNs, which is within the acceptable limits.' and provides running times in Table 4. However, it does not specify any exact hardware details like GPU/CPU models or specific machine configurations used for their experiments. |
| Software Dependencies | No | The paper mentions using the 'Adam optimizer' and refers to the 'UMAP tool' (with a link to its documentation) for visualization, but it does not specify version numbers for any software dependencies, frameworks (like PyTorch or TensorFlow), or libraries. |
| Experiment Setup | Yes | To avoid the time dimensions being 0, we uniformly perturb each dimension with a small random value in the interval [ , ], where = 0.02 in practice. In addition, the same 10-dimensional embedding and 2 hidden layers are used for all baselines to ensure a fair comparison. The learning rate is set to 0.01, the learning rate of curvature is set to 0.0001. Q-GCN is implemented with the Adam optimizer. We repeat the experiments 10 times via different random seeds influencing weight initialization and data batching. To ensure a fair comparison, we set the same 16-dimension hidden embedding, 0.01 initial learning rate and 0.0001 learning rate for curvature. The optimal regularization with weight decay, dropout rate, the number of layers and activation functions are obtained by grid search for each method. |