Aggregation Buffer: Revisiting DropEdge with a New Parameter Block
Authors: Dooho Lee, Myeong Kong, Sagad Hamid, Cheonwoo Lee, Jaemin Yoo
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the effectiveness of AGGB in improving the robustness and overall accuracy of GNNs across 12 node classification benchmarks. In addition, we show that AGGB works as a unifying solution to structural inconsistencies such as degree bias and structural disparity, both of which arise from structural variations in graph datasets. ... We evaluate the accuracy of node classification for 12 widely-used benchmark graphs... All reported performances, including the ablation studies, are averaged over ten independent runs with different random seeds and splits... |
| Researcher Affiliation | Academia | 1School of Electrical Engineering, KAIST, Daejeon, Republic of Korea 2Computer Science Department, University of Münster, Münster, Germany. Correspondence to: Jaemin Yoo <EMAIL>. |
| Pseudocode | No | The paper includes theoretical analysis and mathematical formulations but does not contain a clearly labeled 'Pseudocode' or 'Algorithm' block. |
| Open Source Code | Yes | Code and datasets are available at https: //github.com/dooho00/agg-buffer. |
| Open Datasets | Yes | We evaluate the accuracy of node classification for 12 widely-used benchmark graphs, including Cora, Citeseer, Pubmed, Computers, Photo, CS, Physics, Ogbn-arxiv, Actor, Squirrel and Chameleon (Shchur et al., 2018; Hu et al., 2020a; Pei et al., 2020). For Squirrel and Chameleon, we use the filtered versions provided by (Platonov et al., 2023) via their public repository: https://github.com/yandex-research/ heterophilous-graphs. The remaining 10 datasets are sourced from the Deep Graph Library (DGL) (Wang et al., 2019). |
| Dataset Splits | Yes | We adopt the public dataset splits for Ogbn-arxiv (Hu et al., 2020a), Actor, Squirrel and Chameleon (Pei et al., 2020; Platonov et al., 2023). For the remaining eight datasets, we use an independently randomized 10%/10%/80% split for training, validation, and test, respectively. Our experiments are conducted using a two-layer GCN (Kipf & Welling, 2017), with hyperparameters selected via grid search based on validation accuracy across five runs, following prior works (Luo et al., 2024). |
| Hardware Specification | Yes | All experiments are conducted on an NVIDIA RTX A6000 GPU with 48 GB of memory. |
| Software Dependencies | No | The paper mentions using the Adam optimizer and the Deep Graph Library (DGL) but does not specify their version numbers or other key software dependencies with versions. |
| Experiment Setup | Yes | Our experiments are conducted using a two-layer GCN (Kipf & Welling, 2017), with hyperparameters selected via grid search based on validation accuracy across five runs... The search space included hidden dimensions [64, 256, 512], dropout ratios [0.2, 0.3, 0.5, 0.7], weight decay values [0, 5e-4, 5e-5], and learning rates [1e-2, 1e-3, 5e-3]. We use the Adam optimizer (Kingma & Ba, 2015) for training with early stopping based on validation accuracy, using a patience of 100 epochs across all datasets. ... AGGB requires tuning on three key hyperparameters: the dropout ratio, Drop Edge ratio, and the coefficient λ... λ values: [1, 0.5, 0.1], Drop Edge ratio: [0.2, 0.5, 0.7, 1.0], Dropout ratio: [0, 0.2, 0.5, 0.7]. For hyperparameter tuning, we follow the same process used for training the base GCN, conducting a search across five independent runs and selecting the configuration with the highest validation accuracy. To ensure reproducibility, we provide the detailed hyperparameters for training AGGB across datasets in Table 9. |