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 [1].
Improving Graph Neural Networks on Multi-node Tasks with the Labeling Trick
Authors: Xiyuan Wang, Pan Li, Muhan Zhang
JMLR 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments include various multi-node representation learning tasks: undirected link prediction, directed link prediction, hyperlink prediction, and subgraph prediction. Labeling trick boosts GNNs on all these tasks. All metrics in this section are the higher the better. Datasets are detailed in Appendix C. |
| Researcher Affiliation | Academia | Xiyuan Wang EMAIL Institute for Artificial Intelligence Peking University Beijing, China. Pan Li EMAIL School of Electrical and Computer Engineering Georgia Institute of Technology Atlanta, USA. Muhan Zhang EMAIL Institute for Artificial Intelligence Peking University Beijing, China. |
| Pseudocode | No | The paper describes methods and theoretical proofs but does not include any clearly labeled pseudocode or algorithm blocks with structured steps. |
| Open Source Code | Yes | Our code is available at https://github. com/Graph PKU/Labeling Trick. |
| Open Datasets | Yes | We also use OGB datasets (Hu et al., 2020): ogbl-ppa, ogbl-collab, ogbl-ddi, and ogbl-citation2. Among them, ogbl-ppa is a protein-protein association graph where the task is to predict biologically meaningful associations between proteins. ogbl-collab is an author collaboration graph, where the task is to predict future collaborations. ogbl-ddi is a drug-drug interaction network, where each edge represents an interaction between drugs which indicates the joint effect of taking the two drugs together is considerably different from their independent effects. ogbl-citation2 is a paper citation network, where the task is to predict missing citations. |
| Dataset Splits | Yes | Table 7: Statistics and evaluation metrics of undirected link prediction datasets. ... Split ratio: USAir 0.85/0.05/0.10; NS 0.85/0.05/0.15; PB 0.85/0.05/0.15; Yeast 0.85/0.05/0.15; C.ele 0.85/0.05/0.15; Power 0.85/0.05/0.15; Router 0.85/0.05/0.15; E.coli 0.85/0.05/0.15; ogbl-ppa fixed; ogbl-collab fixed; ogbl-ddi fixed; ogbl-citation2 fixed. |
| Hardware Specification | Yes | All our models run on an Nvidia 3090 GPU on a Linux server. |
| Software Dependencies | No | We leverage Pytorch Geometric and Pytorch for model development. All our models run on an Nvidia 3090 GPU on a Linux server. |
| Experiment Setup | Yes | We use Adam optimizer and constant learning rate for all our models. Main hyperparameters for our models are listed in Table 6. More detailed configuration of each experiments is provided in our code. |