Rethinking Graph Neural Networks From A Geometric Perspective Of Node Features

Authors: Feng Ji, Yanan Zhao, KAI ZHAO, Hanyang Meng, Jielong Yang, Wee Peng Tay

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 5 EXPERIMENTS... The results are shown in Table 2. Among 48 comparisons, the above tricks significantly (in terms of p-value) improve the performance in 34 instances, and the improvement is insignificant in 14 instances, mainly for Chameleon, Squirrel, and Actor datasets.
Researcher Affiliation Academia 1School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore 2School of Internet of Things Engineering, Jiangnan University, Wuxi, China
Pseudocode No The paper describes methods using mathematical formulas and textual explanations rather than structured pseudocode or algorithm blocks.
Open Source Code Yes The source code to reproduce our results can be found at https://github.com/Yanan Zhao0630/M-AE-M-AEN (base model: ACM-GCN).
Open Datasets Yes The following datasets are used and studied at various places in the paper (including the appendices): Cora, Citeseer, Pub Med, Ogbn-arxiv, Texas, Cornell, Wisconsin, Chameleon, Squirrel, Actor, Penn94, ar Xiv-year, and genius (see Lim et al. (2021)).
Dataset Splits Yes Table 6: Dataset statistics... Data splits: standard, 48%/32%/20%, 50%/25%/25%
Hardware Specification Yes Experiments are performed on a workstation with a single NVIDIA Ge Force RTX 3090 GPU and 24GB memory.
Software Dependencies No The paper mentions several GNN models (GCN, GAT, ACM-GCN, Graph CON, CDE, Glo GNN) and states that experiments are performed on a workstation, but does not specify software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions).
Experiment Setup Yes D.2 HYPERPARAMETERS There are two hyperparameters used in the tricks in Section 5: η for the edge addition probability and E for the number of epochs in training. They are tuned according to the following general procedure. For η, we will first consider η = 1, 0.5, 0.2, 0.1, 0.001, 0.0001 and then fine-tune around one of these values using the validation set... For the number of epochs, let E0 be the number of epochs of the base model (E0 = 1000 for CDE and Glo GNN and = 200 for other base models). We consider E = E0/20, E0/10, E0/4, E0/2 and then fine-tune around one of these values.