Polynormer: Polynomial-Expressive Graph Transformer in Linear Time

Authors: Chenhui Deng, Zichao Yue, Zhiru Zhang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Polynormer has been evaluated on 13 homophilic and heterophilic datasets, including large graphs with millions of nodes. Our extensive experiment results show that Polynormer outperforms state-of-the-art GNN and GT baselines on most datasets, even without the use of nonlinear activation functions.
Researcher Affiliation Academia Chenhui Deng, Zichao Yue, Zhiru Zhang Cornell University, Ithaca, USA
Pseudocode Yes Code 1: Py Torch-style Pseudocode for Polynormer
Open Source Code Yes Source code of Polynormer is freely available at: github.com/cornell-zhang/Polynormer.
Open Datasets Yes Appendix E shows the details of all 13 datasets used in our experiments, which consist of 7 homophilic and 6 heterophilic graphs. For Computer, Photo, CS, and Physics datasets, we adhere to the widely accepted practice of randomly dividing nodes into training (60%), validation (20%), and test (20%) sets (Chen et al., 2022b; Shirzad et al., 2023).
Dataset Splits Yes For Computer, Photo, CS, and Physics datasets, we adhere to the widely accepted practice of randomly dividing nodes into training (60%), validation (20%), and test (20%) sets (Chen et al., 2022b; Shirzad et al., 2023).
Hardware Specification Yes We conduct all experiments on a Linux machine equipped with an Intel Xeon Gold 5218 CPU (featuring 8 cores @ 2.30 GHz) and 4 RTX A6000 GPUs (each with 48 GB of memory).
Software Dependencies No The paper mentions 'Py G (Fey & Lenssen, 2019) and DGL (Wang et al., 2019)' as frameworks but does not specify exact version numbers for these or any other software dependencies.
Experiment Setup Yes Detailed hyperparameter settings of baselines and Polynormer are available in Appendix H. In Table 5, we provide the critical hyperparameters of Polynormer used with each dataset.