GATE: How to Keep Out Intrusive Neighbors

Authors: Nimrah Mustafa, Rebekka Burkholz

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

Reproducibility Variable Result LLM Response
Research Type Experimental 5. Experiments We validate the ability of GATE to perform the appropriate amount of neighborhood aggregation, as relevant for the given task and input graph, on both synthetic and real-world graphs.
Researcher Affiliation Academia 1CISPA Helmholtz Center for Information Security, 66123 Saarbr ucken, Germany.
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes Our experimental code is available at https://github.com/Relational ML/GATE.git.
Open Datasets Yes On real-world datasets, GATE performs competitively on homophilic datasets and is substantially better than GAT on heterophilic datasets. Furthermore, up to our knowledge, it achieves a new state of the art on the relatively large OGB-arxiv dataset (i.e., 79.57 0.84% test accuracy). ... We evaluate GATE on relatively large-scale real-world node classification tasks, namely on five heterophilic benchmark datasets (Platonov et al., 2023) (see Table 3) and three OGB datasets (Hu et al., 2021) (see Table 5).
Dataset Splits Yes Nodes are divided randomly into train/validation/test split with a 2 : 1 : 1 ratio. ... Real-world datasets use their standard train/test/validation splits, i.e. those provided by Pytorch Geometric for Planetoid datasets Cora and Citeseer, by OGB framework for OGB datasets, and by (Platonov et al., 2023) for all remaining real-world datasets.
Hardware Specification No The paper does not explicitly describe the specific hardware (e.g., GPU/CPU models, memory) used to run its experiments.
Software Dependencies No The paper mentions Pytorch Geometric and Adam optimizer but does not specify their version numbers or the versions of other key software dependencies.
Experiment Setup Yes For synthetic datasets, the network width is fixed to 64 in all cases. ... For all synthetic data, a learning rate of 0.005 is used. Real-world datasets use their standard train/test/validation splits... the learning rate is adjusted for different real-world datasets to enable stable training of models as specified in Table 6.