Training-free Graph Neural Networks and the Power of Labels as Features

Authors: Ryoma Sato

TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In the experiments, we confirm that TFGNNs outperform existing GNNs in the training-free setting and converge with much fewer training iterations than traditional GNNs.
Researcher Affiliation Academia Ryoma Sato EMAIL National Institute of Informatics
Pseudocode No The paper defines the TFGNN architecture and its initialization using mathematical equations (19-29) rather than a separate pseudocode or algorithm block.
Open Source Code Yes Reproducibility: Our code is available at https://github.com/joisino/laf.
Open Datasets Yes We use the Planetoid datasets (Cora, Cite Seer, Pub Med) [54], Coauthor datasets, and Amazon datasets [42] in the experiments.
Dataset Splits Yes We use 20 nodes per class for training, 500 nodes for validation, and the rest for testing in the Planetoid datasets following Kipf et al. [20], and use 20 nodes per class for training, 30 nodes per class for validation, and the rest for testing in the Coauthor and Amazon datasets following Shchur et al. [42].
Hardware Specification No The paper does not provide specific hardware details such as GPU or CPU models used for the experiments.
Software Dependencies No The paper mentions using Adam W for training but does not provide specific version numbers for any software libraries or frameworks used in the experiments.
Experiment Setup Yes We use three layered models with the hidden dimension 32 unless otherwise specified. We train all the models with Adam W [25] with learning rate 0.0001 and weight decay 0.01.