Distance-Restricted Folklore Weisfeiler-Leman GNNs with Provable Cycle Counting Power

Authors: Junru Zhou, Jiarui Feng, Xiyuan Wang, Muhan Zhang

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on both synthetic datasets and molecular datasets verify our theory. In this section, we empirically evaluate the performance of d-DRFWL(2) GNNs (especially for the case of d = 2) and verify our theoretical results.
Researcher Affiliation Academia Junru Zhou1 Jiarui Feng2 Xiyuan Wang1 Muhan Zhang1 1Institute for Artificial Intelligence, Peking University 2Department of CSE, Washington University in St. Louis
Pseudocode No The paper describes algorithms using mathematical equations and descriptive text, such as equations (1) to (11), but it does not include explicitly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code Yes Our code for all experiments, including those in Section 6 of the main paper and in Appendix F, is available at https://github.com/zml72062/DR-FWL-2.
Open Datasets Yes To answer Q1, we perform node-level substructure counting tasks on the synthetic dataset in [33, 60]. The QM9 and ZINC datasets are provided by PyTorch Geometric package [22]. The ogbg-molhiv and ogbg-molpcba datasets are provided by Open Graph Benchmark (OGB) [31].
Dataset Splits Yes The synthetic dataset contains 5,000 random graphs, and the training/validation/test splitting is 0.3/0.2/0.5. The training/validation/test splitting for QM9 is 0.8/0.1/0.1. The training/validation/test splittings for ZINC, ogbg-molhiv and ogbg-molpcba are provided in the original releases.
Hardware Specification No The paper mentions 'maximal GPU memory usage' and lists 'Memory (GB)' in tables (e.g., Table 4, 14), implying the use of GPUs, but it does not specify any particular GPU models (e.g., NVIDIA A100, RTX 2080 Ti), CPU models, or other detailed hardware specifications used for experiments.
Software Dependencies No The paper mentions 'PyTorch Geometric package [22]' without specifying a version number. No other software dependencies are listed with specific version numbers, which is necessary for reproducibility.
Experiment Setup Yes We use Adam optimizer with initial learning rate 0.001, and use plateau scheduler with patience 10, decay factor 0.9 and minimum learning rate 10 5. We train our model for 2,000 epochs. The batch size is 256.