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].

Line Graph Vietoris-Rips Persistence Diagram for Topological Graph Representation Learning

Authors: Jaesun Shin, Eunjoo Jeon, Taewon Cho, Namkyeong Cho, Youngjune Gwon

JMLR 2024 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally we empirically validate superior performance of our models on several graph classification and regression benchmarks. Keywords: Graph Neural Network, Persistence Diagram, Topological Data Analysis, Weisfeiler-Lehman Test, Vietoris-Rips Filtration
Researcher Affiliation Collaboration Jaesun Shin EMAIL Samsung SDS; Namkyeong Cho1 EMAIL Center for Mathematical Machine Learning and its Applications(CM2LA), Department of Mathematics POSTECH
Pseudocode Yes Pseudocode for the construction of LGVR is described in Algorithm 1 (See Figure 4 for the overall framework).
Open Source Code Yes Our code is publicly available at https://github.com/samsungsds-research-papers/LGVR.
Open Datasets Yes For classification, we test our method on 7 benchmark graph datasets: 5 bioinformatics datasets (MUTAG, PTC, PROTEINS, NCI1, NCI109) that represent chemical compounds or protein substructures, and two social network datasets (IMDB-B, IMDB-M) (Yanardag and Vishwanathan (2015)). For the regression task, we experiment on a standard graph benchmark QM9 dataset (Ramakrishnan et al. (2014); Ruddigkeit et al. (2012); Wu et al. (2018)).
Dataset Splits Yes Since there is no separate test set for these datasets, for a fair comparison, we follow the standard 10-fold cross-validation based on the same split used by Zhang et al. (2018) and report the results according to the evaluation protocol described by Maron et al. (2019): ... The dataset is split into 80% train, 10% validation, and 10% test.
Hardware Specification Yes We run all experiments on a single DGX-A100 GPU.
Software Dependencies No The paper mentions using the pytorch-geometric library and Adam optimizer but does not specify version numbers for these software components. For example, in Section D.1.3, it states: "Each graph is represented by an adjacency matrix and input node features, which can be obtained from the pytorch-geometric library Fey and Lenssen (2019)." However, no version for pytorch-geometric is provided.
Experiment Setup Yes All hyperparameters were carried out on the same set based on C (Appendix D.2): the learning rate is set to {5e-3, 1e-3, 5e-4, 1e-4, 5e-5} for GIN while it is set to {1e-4, 5e-5} for PPGN. The decay rate is set to {0.5, 0.75, 1.0} with Adam optimizer (Kingma and Ba (2014)), and we implement all the models by tuning hyperparameters based on the validation score. Tables 7 and 8 in Appendix D.2 provide specific LR, DR, BS, Ep for each dataset.