GraphFM: Improving Large-Scale GNN Training via Feature Momentum

Authors: Haiyang Yu, Limei Wang, Bokun Wang, Meng Liu, Tianbao Yang, Shuiwang Ji

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, we observe that Graph FM-IB can effectively alleviate the neighborhood explosion problem of existing methods. In addition, Graph FM-OB achieves promising performance on multiple large-scale graph datasets.
Researcher Affiliation Academia 1Department of Computer Science & Engineering, Texas A&M University, TX, USA 2Department of Computer Science, The University of Iowa, IA, USA.
Pseudocode Yes Algorithm 1 Graph FM-IB and Algorithm 2 Graph FM-OB are provided in the paper.
Open Source Code Yes Our code is implemented in the DIG (Dive into Graphs) library (Liu et al., 2021), which is a turnkey library for graph deep learning research and publicly available1. 1https://github.com/divelab/DIG/tree/dig/dig/lsgraph
Open Datasets Yes We evaluate our proposed algorithms Graph FM-IB and Graph FM-OB with extensive experiments on the node classification task on five large-scale graphs, including Flickr (Zeng et al., 2019), Yelp (Zeng et al., 2019), Reddit (Hamilton et al., 2017), ogbn-arxiv (Hu et al., 2021) and ogbn-products (Hu et al., 2021).
Dataset Splits Yes Table 1: Statistics and properties of the datasets. The m denotes the multi-label classification task, and s denotes single label classification task. Dataset # of nodes # of edges Avg. degree # of features # of classes Train/Val/Test ... Flickr ... 0.500/0.250/0.250
Hardware Specification Yes In addition, we conduct our experiments on Nvidia Ge Force RTX 2080 with 11GB memory, and Intel Xeon Gold 6248 CPU.
Software Dependencies No The paper states: 'The implementation of our methods is based on the Py Torch (Paszke et al., 2019), and Pytorch_geometric (Fey & Lenssen, 2019).' While PyTorch has a year cited, specific version numbers for both key software components (PyTorch and Pytorch_geometric) are not provided in the text.
Experiment Setup Yes We explore the feature momentum hyper-parameter β in the range from 0.1 to 0.9. We select the learning rate from {0.01, 0.05, 0.001} and dropout from {0.0, 0.1, 0.3, 0.5}. Due to the over-fitting problem on the ogbn-products dataset, we set the edge drop (Rong et al., 2019) ratio at 0.8 during training for this particular dataset.