Convolutional Neural Networks on Graphs with Chebyshev Approximation, Revisited
Authors: Mingguo He, Zhewei Wei, Ji-Rong Wen
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conducted an extensive experimental study to demonstrate that Cheb Net II can learn arbitrary graph convolutions and achieve superior performance in both fulland semisupervised node classification tasks. Most notably, we scale Cheb Net II to a billion graph ogbn-papers100M, showing that spectral-based GNNs have superior performance. Our code is available at https://github.com/ivam-he/Cheb Net II. |
| Researcher Affiliation | Academia | Mingguo He Renmin University of China mingguo@ruc.edu.cn Zhewei Wei Renmin University of China zhewei@ruc.edu.cn Ji-Rong Wen Renmin University of China jrwen@ruc.edu.cn |
| Pseudocode | No | The paper describes mathematical formulations and processes but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | Yes | Our code is available at https://github.com/ivam-he/Cheb Net II. |
| Open Datasets | Yes | We evaluate Cheb Net II on several real-world graphs for the Semiand Full-supervised node classification tasks. The datasets include three homophilic citation graphs: Cora, Citeseer, and Pubmed [35, 45], five heterophilic graphs: the Wikipedia graphs Chameleon and Squirrel [34], the Actor co-occurrence graph, and webpage graphs Texas and Cornell from Web KB* [30], two large citation graphs: ogbn-arxiv and ogbn-papers100M [18], as well as six large heterophilic graphs: Penn94, pokec, ar Xiv-year, genius, twitch-gamers and wiki [25]. |
| Dataset Splits | Yes | Specifically, we apply the standard training/validation/testing split [45] on the three homophilic citation datasets (i.e., Cora, Citeseer, and Pubmed), with 20 nodes per class for training, 500 nodes for validation, and 1,000 nodes for testing. Since this standard split can not be used for very small graphs (e.g. Texas), we use the sparse splitting [6] with the training/validation/test sets accounting for 2.5%/2.5%/95%, respectively, on the five heterophilic datasets. |
| Hardware Specification | Yes | All the experiments are carried out on a machine with an NVIDIA RTX8000 GPU (48GB memory), Intel Xeon CPU (2.20 GHz) with 40 cores, and 512 GB of RAM. |
| Software Dependencies | No | The paper mentions using 'Pytorch Geometric library implementations [11]' but does not specify version numbers for PyTorch, Python, or other software dependencies. |
| Experiment Setup | Yes | For Cheb Net II, we use Equation (8) as the propagation process and use the Re Lu function to reparametrize γj, maintaining the non-negativity of the filters [17]. We set the hidden units as 64 and K = 10 for the all datasets as the same as GPR-GNN [6] and Bern Net [17]. We employ the Adam SGD optimizer [20] with an early stopping of 200 and a maximum of 1000 epochs to train Cheb Net II. |