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..
Can Graph Neural Networks Count Substructures?
Authors: Zhengdao Chen, Lei Chen, Soledad Villar, Joan Bruna
NeurIPS 2020 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We then conduct experiments that support the theoretical results for MPNNs and 2-IGNs. We empirically demonstrate that it can count both subgraphs and induced subgraphs on random synthetic graphs while also achieving competitive performances on molecular datasets. |
| Researcher Affiliation | Academia | Zhengdao Chen New York University EMAIL Lei Chen New York University EMAIL Soledad Villar Johns Hopkins University EMAIL Joan Bruna New York University EMAIL |
| Pseudocode | No | The paper describes the models and methods using mathematical equations and textual explanations but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | 1Code available at https://github.com/leichen2018/GNN-Substructure-Counting. |
| Open Datasets | Yes | We evaluate LRP on the molecular prediction datasets ogbg-molhiv [63], QM9 [54] and ZINC [14]. |
| Dataset Splits | Yes | For the counting tasks, we generate a dataset of 1000 graphs, consisting of 800 training graphs, 100 validation graphs, and 100 testing graphs, all of them of size ranging from 10 to 50 nodes. For ogbg-molhiv, QM9 and ZINC, we follow the settings described in [63, 39, 14] for data splits... |
| Hardware Specification | Yes | Each model is trained on 1080ti five times with different random seeds. |
| Software Dependencies | No | The paper mentions various GNN models and general software components but does not provide specific version numbers for any libraries, frameworks, or solvers (e.g., Python, PyTorch, TensorFlow, CUDA versions). |
| Experiment Setup | Yes | We use Adam optimizer [30] with a learning rate of 0.001 and a weight decay of 10−5... For ogbg-molhiv, the total number of epochs is set to 1000 with a batch size of 32. For QM9, the total number of epochs is set to 2000 with a batch size of 128. For ZINC, the total number of epochs is set to 2000 with a batch size of 128. |