LazyGNN: Large-Scale Graph Neural Networks via Lazy Propagation

Authors: Rui Xue, Haoyu Han, Mohamadali Torkamani, Jian Pei, Xiaorui Liu

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

Reproducibility Variable Result LLM Response
Research Type Experimental Comprehensive experiments demonstrate its superior prediction performance and scalability on large-scale benchmarks. The implementation of Lazy GNN is available at https: //github.com/RXPHD/Lazy_GNN.
Researcher Affiliation Collaboration 1North Carolina State University, Raleigh, US 2Michigan State University, East Lansing, US 3Amazon, US (this work does not relate to the author s position at Amazon) 4Duke University, Durham, US.
Pseudocode No The paper includes conceptual diagrams (Figure 3 and Figure 4) but no structured pseudocode or algorithm blocks.
Open Source Code Yes The implementation of Lazy GNN is available at https: //github.com/RXPHD/Lazy_GNN.
Open Datasets Yes We conduct experiments on multiple large-scale graph datasets including REDDIT, YELP, FLICKR, ogbn-arxiv, and ogbn-products (Hu et al., 2020).
Dataset Splits Yes We conduct experiments on multiple large-scale graph datasets including REDDIT, YELP, FLICKR, ogbn-arxiv, and ogbn-products (Hu et al., 2020). The hyperparameter tuning of baselines closely follows the setting in GNNAuto Scale (Fey et al., 2021). The convergence of validation accuracy in Figure 5 demonstrates that Lazy GNN has a comparable convergence speed with GCN (GAS) and GCNII (GAS), and is slightly faster than APPNP (GAS) in terms of the number of training epochs.
Hardware Specification No The paper mentions running experiments on CPU and GPU memory but does not specify particular models, types, or configurations of the hardware used.
Software Dependencies No The paper does not provide specific version numbers for any software dependencies or libraries used.
Experiment Setup Yes For Lazy GNN, hyperparameters are tuned from the following search space: (1) learning rate: {0.01, 0.001, 0.0001}; (2) weight decay: {0, 5e 4, 5e 5}; (3) dropout: {0.1, 0.3, 0.5, 0.7}; (4) propagation layers : L {1, 2}; (5) MLP layers: {3, 4}; (6) MLP hidden units: {256, 512}; (7) α {0.01, 0.1, 0.2, 0.5, 0.8}; (8) β and γ are simply set as 0.5 in most cases, but a further tuning can improve the performance.