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..
Subgraph Neural Networks
Authors: Emily Alsentzer, Samuel Finlayson, Michelle Li, Marinka Zitnik
NeurIPS 2020 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical results for subgraph classification on eight datasets show that SUBGNN achieves considerable performance gains, outperforming strong baseline methods, including node-level and graph-level methods, by 19.8% over the strongest baseline. |
| Researcher Affiliation | Academia | Emily Alsentzer Harvard University, MIT EMAIL Samuel G. Finlayson Harvard University, MIT EMAIL Michelle M. Li Harvard University EMAIL Marinka Zitnik Harvard University EMAIL |
| Pseudocode | Yes | A full description of SUBGNN is detailed in Appendix A. (referring to Algorithm 1 in Appendix A.1) |
| Open Source Code | Yes | Code and datasets are available at https://github.com/mims-harvard/SubGNN. |
| Open Datasets | Yes | Code and datasets are available at https://github.com/mims-harvard/SubGNN. |
| Dataset Splits | No | The paper mentions using 'validation loss' for early stopping in Appendix D, implying the existence of a validation set, but it does not specify the exact percentages or counts for train/validation/test splits. |
| Hardware Specification | Yes | With pre-computation, training SUB-GNN on the real-world dataset PPI-BP takes 30s per epoch on a single GTX 1080 GPU |
| Software Dependencies | No | We use PyTorch Lightning [17] for modularity and reproducibility. All experiments were implemented in Python 3.7.4. (Missing versions for key libraries like PyTorch Lightning or PyTorch itself.) |
| Experiment Setup | Yes | We train all models for 100 epochs, with early stopping enabled based on validation loss with a patience of 10. We use the Adam optimizer with default parameters (β1 = 0.9, β2 = 0.999, = 1e−8). The learning rate is initialized to 1e−3 and reduced by a factor of 0.5 every 25 epochs. We perform a random search for batch sizes from {1, 2, 4, 8, 16, 32, 64, 128} and report the best results. |