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 Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Relieving the Over-Aggregating Effect in Graph Transformers
Authors: Junshu Sun, Wanxing Chang, Chenxue Yang, Qingming Huang, Shuhui Wang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Evaluations show that Wideformer can effectively mitigate over-aggregating. As a result, the backbone methods can focus on the informative messages, achieving superior performance compared to baseline methods. We demonstrate the effectiveness of Wideformer on thirteen real-world datasets, which consistently relieves over-aggregating and benefits backbones to achieve superior performance. |
| Researcher Affiliation | Collaboration | 1State Key Lab. of AI Safety, ICT, CAS 2University of Chinese Academy of Sciences 3DAMO Academy, Alibaba Group 4Agriculture Information Institute, CAAS EMAIL EMAIL EMAIL EMAIL |
| Pseudocode | Yes | Algorithm 1 Center-selection Function Cluster |
| Open Source Code | Yes | Codes are available at https://github.com/sunjss/over-aggregating. |
| Open Datasets | Yes | We adopt thirteen real-world datasets, including both heterophilic and homophilic graphs. For baseline methods, both GNNs and graph transformers are adopted. Please refer to Appendix D.2 for detailed experimental setups. The attention entropy experiments in Fig. 1, Fig. 3, and Fig. 5 are conducted on Cora, Amazon Photo, Amazon Computers, Coauthor CS, Coauthor Physics, tolokers, amazon-ratings, minesweeper, and ogb-arxiv. The cluster ablation study in Fig. 7 is conducted and averaged on all the datasets included in this paper. |
| Dataset Splits | No | The paper does not explicitly provide specific training/test/validation dataset splits (e.g., percentages, sample counts, or explicit references to predefined splits from citations for *their* experiments). It mentions dataset names which might have standard splits, but this is not explicitly stated in the paper's experimental setup. |
| Hardware Specification | Yes | The framework is implemented with Py Torch [31] and Py Torch Geometric [12], and trained on a single NVIDIA A100. |
| Software Dependencies | No | The framework is implemented with Py Torch [31] and Py Torch Geometric [12], and trained on a single NVIDIA A100. (Specific version numbers for PyTorch or PyTorch Geometric are not provided.) |
| Experiment Setup | Yes | We perform grid search based on the validation performance of the models as follows: Graph GPS. We search the number of graph transformer layers in {1, , 6}, the number of hidden dimensions in {64, 80, 128, 256}, the number of heads in {1, 2, 4}, dropout in {0.1, 0.2, 0.3, 0.5}, and learning rate in {5e 4, 1e 3, 1e 2}. The rest hyperparameters are fixed as in the original implementation. SGFormer. The GNN backbone is implemented as GCN [17]. The number of GCN layers is searched in {1, , 10}, the number of hidden dimensions in {64, 80, 128, 256}, the number of heads in {1, 2, 4}, dropout in {0.1, 0.2, 0.3, 0.5}, and learning rate in {1e 3, 5e 3, 1e 2}. The rest hyperparameters are fixed as in the original implementation. Polynormer. The GNN backbone is implemented as GAT [48]. We search the number of graph transformer layers in {1, , 6}, the number of GAT layers in {1, , 10}, the number of hidden dimensions in {64, 80, 128, 256}, the number of heads in {1, 2, 4, 8}, dropout in {0.1, 0.2, 0.3, 0.5}, and learning rate in {5e 4, 1e 3}. The rest hyperparameters are fixed following the original implementation. |