Weisfeiler and Lehman Go Paths: Learning Topological Features via Path Complexes

Authors: Quang Truong, Peter Chin

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical validation of our assertions is offered through evaluations of PCN on various real-world benchmarks and a collection of 9 strongly regular graph (SRGs) families. Notably, our proposed approach achieves superior performance without the need for assumptions about graph sub-structures other than paths, which are inherent in every connected graph.
Researcher Affiliation Academia Thayer School of Engineering, Dartmouth College cong.minh.quang.truong.th@dartmouth.edu, peter.chin@dartmouth.edu
Pseudocode No The paper describes the PWL procedure and message-passing updates in narrative text and equations, but does not provide a formally structured pseudocode or algorithm block.
Open Source Code No Please refer to the Appendix of the extended version for the detailed proofs, formulae, experiments, computational analysis, and other resources. https://arxiv.org/abs/2308.06838
Open Datasets Yes TUDataset benchmarks, encompassing a broad spectrum of graph datasets from biology, chemistry, and social networks, are proposed in (Morris et al. 2020). We also evaluate our model on the ZINC dataset, a common graph benchmark for regression tasks (Sterling and Irwin 2015; Dwivedi et al. 2023). We also evaluate our model on the OGBG-MOLHIV dataset (Hu et al. 2020), which contains about 41k graphs for the graph binary classification task.
Dataset Splits Yes Specifically, we report the highest mean test accuracy across a 10-fold cross-validation as indicated in (Xu et al. 2019). Hu et al. applies a scaffold splitting procedure (Wu et al. 2018) for the OGBG-MOLHIV dataset; thus we report the performance of our model on both the validation and test sets in Table 2.
Hardware Specification No The paper does not explicitly describe the specific hardware used (e.g., GPU/CPU models, memory amounts) for running its experiments.
Software Dependencies No The paper mentions several software tools and libraries (e.g., PyTorch Geometric, Weights and Biases, PyTorch, NetworkX, graph-tool), but does not provide specific version numbers for them.
Experiment Setup No Detailed hyperparameter settings along with relevant ablation studies are documented in the Appendix. These details are not present in the provided main text of the paper.