Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Node-Specific Space Selection via Localized Geometric Hyperbolicity in Graph Neural Networks
Authors: See Hian Lee, Feng Ji, Wee Peng Tay
TMLR 2024 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our model on both node classification and link prediction tasks and observe promising performance compared to baseline models. ... In this section, we evaluate JSGNN on node classification (NC) and link prediction (LP) tasks. Dataset statistics, model settings, and model size and complexity are discussed in Appendix A and Appendix B. ... Our empirical results demonstrate that JSGNN frequently outperforms the baselines, especially HGAT and GAT which are the building blocks of JSGNN. |
| Researcher Affiliation | Academia | See Hian Lee EMAIL Department of Electrical and Electronic Engineering Nanyang Technological University. Ji Feng EMAIL Department of Electrical and Electronic Engineering Nanyang Technological University. Tay Wee Peng EMAIL Department of Electrical and Electronic Engineering Nanyang Technological University. |
| Pseudocode | No | The paper describes the model architecture and equations but does not present a clearly labeled pseudocode or algorithm block. |
| Open Source Code | Yes | The source code can be found at https://github.com/amblee0306/JSGNN_Mixed_Space_GNN. |
| Open Datasets | Yes | A total of seven benchmark datasets are employed for both NC and LP. Specifically, three citation datasets: Cora, Citeseer, Pubmed; a flight network: Airport; a disease propagation tree: Disease; an Amazon co-purchase graph dataset: Photo; and a coauthor dataset: CS. |
| Dataset Splits | Yes | In the first split, we followed the standard split for semi-supervised settings used in Kipf & Welling (2016); Veličković et al. (2018); Monti et al. (2017); Chamberlain et al. (2021b); Zhu et al. (2020); Chamberlain et al. (2021a); Hamilton et al. (2017); Liu et al. (2022b); Feng et al. (2020). The train set consists of 20 train examples per class while the validation set and test set consist of 500 samples and 1,000 samples, respectively. Meanwhile, in the second split, all labels are utilized and the percentages of training, validation, and test sets are set as 60/20/20%. For the Photo and CS datasets, the labeled nodes are also split into three sets where 60% of the nodes made up the training set, and the rest of the nodes were divided equally to form the validation and test sets. Airport and Disease datasets were split in similar settings as Zhu et al. (2020). |
| Hardware Specification | No | The paper discusses runtime and scalability but does not provide specific hardware details like GPU/CPU models or memory amounts used for experiments. |
| Software Dependencies | No | The implementation is based on the setup in https://github.com/Cherise Zhu/GIL. However, specific versions of software dependencies are not mentioned. |
| Experiment Setup | Yes | For all models, the hidden units are set to 16. We set the early stopping patience to 100 epochs with a maximum limit of 2000 epochs. The hyperparameter settings for the baselines are the same as Zhu et al. (2020) if given. ... For JSGNN, we perform a grid search on the following search spaces: Learning rate: [0.01, 0.005]; Dropout probability: [0.0, 0.1, 0.5, 0.6]; Number of layers: [1, 2, 3]; ωnu and ωwas: [1.0, 0.5, 0.2, 0.1, 0.01, 0.005]; q (cf. (11)): [16, 32, 64]. |