Understanding over-squashing and bottlenecks on graphs via curvature
Authors: Jake Topping, Francesco Di Giovanni, Benjamin Paul Chamberlain, Xiaowen Dong, Michael M. Bronstein
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5 EXPERIMENTAL RESULTS |
| Researcher Affiliation | Collaboration | Jake Topping12 , Francesco Di Giovanni3 , Benjamin P. Chamberlain3, Xiaowen Dong1, and Michael M. Bronstein23 1University of Oxford 2Imperial College London 3Twitter |
| Pseudocode | Yes | Algorithm 1: Stochastic Discrete Ricci Flow (SDRF) |
| Open Source Code | No | The paper does not include a direct statement or link for the open-source code for the methodology described. |
| Open Datasets | Yes | We evaluate the methods on nine datasets: Cornell, Texas and Wisconsin from the Web KB dataset4; Chameleon and Squirrel (Rozemberczki et al., 2021) along with Actor (Tang et al., 2009); and Cora (Mc Callum et al., 2000), Citeseer (Sen et al., 2008) and Pubmed (Namata et al., 2012). As in Klicpera et al. (2019), for all experiments we consider the largest connected component of the graph. Footnote 4: http://www.cs.cmu.edu/afs/cs.cmu.edu/project/theo-11/www/wwkb/ |
| Dataset Splits | Yes | When splitting the data into train/validation/test sets, we first separate the data into a development set and the test set... For each of the 100 random splits the development set is divided randomly into a train set and a validation set... For Cora, Citeseer and Pubmed the development set contains 1500 nodes with the rest kept for the test set, and for each random split the train set is chosen to contain 20 nodes of each class while the rest form the validation set. ... For the remaining datasets we perform a 60/20/20 split, with 20% of nodes set aside as the test set and then for each random split the remaining 80% is split into 60% training, 20% validation. |
| Hardware Specification | Yes | Our experiments were performed on a server with the following specifications: Architecture x86_64 CPU 40x Intel(R) Xeon(R) Silver 4210R CPU @ 2.40GHz GPU 4x Ge Force RTX 3090 (24268Mi B/GPU) RAM 126GB OS Ubuntu 20.04.2 LTS |
| Software Dependencies | No | The paper mentions its base model is a GCN, but does not specify software dependencies with version numbers (e.g., Python, PyTorch, or TensorFlow versions). |
| Experiment Setup | Yes | F.4 HYPERPARAMETERS Table 4: Hyperparameters for GCN with no preprocessing (None). Dataset Dropout Hidden depth Hidden dimension Learning rate Weight decay Cornell 0.3060 1 128 0.0082 0.1570 Texas 0.2346 1 128 0.0072 0.0037 Wisconsin 0.2869 1 64 0.0281 0.0113... |