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..
A Signed Graph Approach to Understanding and Mitigating Oversmoothing
Authors: Jiaqi Wang, Xinyi Wu, James Cheng, Yifei Wang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on nine benchmarks across both homophilic and heterophilic settings demonstrate that SBP consistently improves classification accuracy and mitigates oversmoothing, even at depths of up to 300 layers. |
| Researcher Affiliation | Academia | 1 The Chinese University of Hong Kong 2MIT IDSS & LIDS 3MIT CSAIL EMAIL EMAIL |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. While 'Algorithm 20' is mentioned in Appendix C, it refers to an equation (Equation 20) and is not presented as a standalone pseudocode or algorithm block with structured steps. |
| Open Source Code | Yes | Our code is available at https://github.com/kokolerk/sbp. |
| Open Datasets | Yes | We use nine widely-used node classification benchmark datasets (Table 7), where four of them are heterophilic (Texas, Wisconsin, Cornell, Squirrel, and Amazon-rating [36]), and the remaining four are homophilic (Cora [37], Citeseer [38], and Pub Med [39]), including one large-scale dataset (ogbn-arxiv [40]). |
| Dataset Splits | Yes | We use the default splits in torch_geometric. |
| Hardware Specification | Yes | We use Tesla-V100-SXM2-32GB in all experiments. |
| Software Dependencies | Yes | We implement all experiments in Python 3.9 with Py Torch Geometric on one NVIDIA Tesla V100 GPU. |
| Experiment Setup | Yes | For SBP, we select the optimal value of λ from the set {0.1, 0.5, 0.9}, fix α = 1, and then choose the best value for β from {0.1, 0.5, 0.9}. ... We fix the learning rate and weight decay in the same dataset and run 100 epochs for every dataset. For the GCN backbone, we follow the [16] settings where we run 5 runs from the seed {0, 1, 2, 3, 4} and calculate the mean and the standard deviation. We fix the hidden dimension to 32 and dropout rate to 0.6. Furthermore, we fix the learning rate to be 0.005 and weight decay to be 5e 4 and run 200 epochs for every dataset. |