Beyond Random Masking: When Dropout meets Graph Convolutional Networks
Authors: Yuankai Luo, Xiao-Ming Wu, Hao Zhu
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our theoretical findings are validated through extensive experiments on both node-level and graph-level tasks across 14 datasets. Notably, GCN with dropout and batch normalization outperforms state-of-the-art methods on several benchmarks, demonstrating the practical impact of our theoretical insights. |
| Researcher Affiliation | Collaboration | Yuankai Luo1,2 & Xiao-Ming Wu2 & Hao Zhu3, 1Beihang University, Beijing, China 2The Hong Kong Polytechnic University, Hong Kong 3Data61r CSIRO, Sydney, Australia |
| Pseudocode | No | The paper contains mathematical formulas, theorems, and proofs within the 'Theoretical Framework' section, but no structured pseudocode or algorithm blocks are explicitly presented. |
| Open Source Code | Yes | Our code is available at https://github. com/LUOyk1999/dropout-theory. |
| Open Datasets | Yes | For node-level tasks, we used 10 datasets: Cora, Cite Seer, Pub Med (Sen et al., 2008), ogbn-arxiv, ogbn-products (Hu et al., 2020), Amazon-Computer, Amazon-Photo, Coauthor-CS, Coauthor-Physics (Shchur et al., 2018), and Wiki CS (Mernyei & Cangea, 2020)... For graph-level tasks, we used MNIST, CIFAR10 (Dwivedi et al., 2023), and two Peptides datasets (functional and structural) (Dwivedi et al., 2022). |
| Dataset Splits | Yes | Cora, Cite Seer, and Pub Med are citation networks, evaluated using the semi-supervised setting and data splits from Kipf & Welling (2017). We used the standard 60%/20%/20% training/validation/test splits and accuracy as the evaluation metric (Chen et al., 2022; Shirzad et al., 2023; Deng et al., 2024). For Wiki CS, we adopted the official splits and metrics (Mernyei & Cangea, 2020). For large-scale graphs, we included ogbn-arxiv and ogbn-products with 0.16M to 2.4M nodes, using OGB s standard evaluation settings (Hu et al., 2020). |
| Hardware Specification | Yes | The experiments are conducted on a single workstation with 8 RTX 3090 GPUs. |
| Software Dependencies | No | The paper mentions 'Py Torch Geometric library' but does not specify its version or any other software dependencies with version numbers. |
| Experiment Setup | Yes | For node-level tasks, we adhered to the training protocols specified in (Deng et al., 2024; Luo et al., 2024b;a), employing BN and adjusting the dropout rate between 0.1 and 0.7. In graph-level tasks, we adopted the settings from (T onshoff et al., 2023; Luo et al., 2025), utilizing BN with a consistent dropout rate of 0.2. All experiments were run with 5 different random seeds, and we report the mean accuracy and standard deviation. |