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].
Geodesic Graph Neural Network for Efficient Graph Representation Learning
Authors: Lecheng Kong, Yixin Chen, Muhan Zhang
NeurIPS 2022 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We present experimental results to show that GDGNN achieves highly competitive performance with SOTA GNN models on various graph learning tasks while taking significantly less time. |
| Researcher Affiliation | Academia | Lecheng Kong Washington University in St. Louis EMAIL Chen Washington University in St. Louis EMAIL Zhang Peking University EMAIL |
| Pseudocode | No | The paper does not include any pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code and data of GDGNN can be found at https://github.com/woodcutter1998/gdgnn. |
| Open Datasets | Yes | We use several types of datasets for link prediction. (1) Knowledge Graph (KG) inductive link prediction datasets. ... We follow the standard inductive split of WN18RR[10] and FB15K237[38] as in Teru et al. [36]. (2) OGB large-scale link prediction dataset [18], including OGBL-COLLAB and OGBL-PPA. We use the official data split. (1) TU datasets contain D&D[11], MUTAG[9], PROTEINS[11], PTC_MR[37]. ... (2) OGB datasets[18], including OGBG-MOLHIV and OGBG-MOLPCBA. (3) Synthetic datasets, including EXP[1] and CSL[28]. We use airport datasets, Brazil-Airport, Europe-Airport, and USA-Airport for our node classification experiments. Following the setting in Li et al. [25], we split the dataset with a train/test/valid ratio of 8:1:1 |
| Dataset Splits | Yes | We follow the standard inductive split of WN18RR[10] and FB15K237[38] as in Teru et al. [36]. We use the official data split. ... following the setting in Li et al. [25], we split the dataset with a train/test/valid ratio of 8:1:1 |
| Hardware Specification | Yes | To make a fair running time comparison, we run all models on 32 CPUs and 1 Nvidia Ge Force 1080Ti GPU. |
| Software Dependencies | No | The paper does not list specific software dependencies with version numbers. |
| Experiment Setup | Yes | For the KG datasets, we search the number of GNN layers in {2, 3, 4, 5} and use 64 as hidden dimensions, and the max search distance for geodesic, dmax, is the same as the number of GNN layers. For the OGB datasets, we search the number of GNN layers in {2, 3, 4} and use 100 as the hidden dimension. We train 50 epochs with a batch size of 64 for the KG datasets, and we train 25 epochs with a batch size of 2048 for the OGB datasets. |