A Metadata-Driven Approach to Understand Graph Neural Networks

Authors: Ting Wei Li, Qiaozhu Mei, Jiaqi Ma

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

Reproducibility Variable Result LLM Response
Research Type Experimental We perform a multivariate sparse regression analysis on the metadata derived from benchmarking GNN performance across diverse datasets, yielding a set of salient data properties. To validate the effectiveness of our data-driven approach, we focus on one identified data property, the degree distribution, and investigate how this property influences GNN performance through theoretical analysis and controlled experiments.
Researcher Affiliation Academia Ting Wei Li University of Michigan tingwl@umich.edu Qiaozhu Mei University of Michigan qmei@umich.edu Jiaqi Ma University of Illinois Urbana-Champaign jiaqima@illinois.edu
Pseudocode No The paper describes methods and processes in text and mathematical formulations but does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code No The paper does not contain any explicit statements or links indicating that the authors have made the source code for their proposed methodology publicly available.
Open Datasets Yes We obtain both the benchmark datasets and the model performance using the Graph Learning Indexer (GLI) library [27]. We include the following benchmark datasets in our regression analysis: cora [42], citeseer [42], pubmed [42], texas [33], cornell [33], wisconsin [33], actor [33], squirrel [33], chameleon [33], arxiv-year [23], snap-patents [23], penn94 [23], pokec [23], genius [23], and twitch-gamers [23].
Dataset Splits Yes We randomly split the nodes into training, validation, and test sets with a ratio of 3:1:1.
Hardware Specification No The paper does not provide specific details regarding the hardware (e.g., specific GPU/CPU models, memory, or cloud instances) used for running the experiments.
Software Dependencies No The paper mentions using an 'R package, MSGLasso [21]', Adam [18], and Adam W [26] as optimizers, but does not provide specific version numbers for these software components or for broader frameworks like Python, PyTorch, or TensorFlow.
Experiment Setup Yes Specifically, we set learning rate = 0.01, weight decay = 0.001, dropout rate = 0.6, max epoch = 10000, and batch size = 256. We use Adam [18] as an optimizer for all models except LINKX. Adam W [26] is used with LINKX in order to comply with Lim et al. [23].