Specformer: Spectral Graph Neural Networks Meet Transformers

Authors: Deyu Bo, Chuan Shi, Lele Wang, Renjie Liao

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

Reproducibility Variable Result LLM Response
Research Type Experimental On synthetic datasets, we show that our Specformer can better recover ground-truth spectral filters than other spectral GNNs. Extensive experiments of both node-level and graph-level tasks on real-world graph datasets show that our Specformer outperforms state-ofthe-art GNNs and learns meaningful spectrum patterns.
Researcher Affiliation Academia Deyu Bo1, Chuan Shi1 , Lele Wang2, Renjie Liao2 Beijing University of Posts and Telecommunications1, University of British Columbia2 {bodeyu, shichuan}@bupt.edu.cn, {lelewang, rjliao}@ece.ubc.ca
Pseudocode No The paper describes the Specformer architecture and its components (e.g., eigenvalue encoding, decoding, graph convolution) in detail, but it does not provide an explicitly labeled 'Pseudocode' or 'Algorithm' block.
Open Source Code Yes Code and data are available at https://github.com/bdy9527/Specformer.
Open Datasets Yes Mol HIV and Mol PCBA are taken from the Open Graph Benchmark (OGB) datasets (Hu et al., 2020). PCQM4Mv2 is a large-scale graph regression dataset (Hu et al., 2021).
Dataset Splits Yes For all datasets, we use the full-supervised split, i.e., 60% for training, 20% for validation, and 20% for testing, as suggested in (He et al., 2021).
Hardware Specification Yes CPU information: Intel(R) Xeon(R) Silver 4210 CPU @ 2.20GHz Ge Force RTX 3090 (24GB)
Software Dependencies Yes Operating system: Linux version 3.10.0-693.el7.x86 64
Experiment Setup Yes Hyperparameters. The hyperparameters of specformer can be seen in Tables 7 and 8. For the node classification task, we use the Adam (Kingma & Ba, 2015) optimizer... For graph-level tasks, we use the Adam W (Loshchilov & Hutter, 2019) optimizer, with the default parameters of ϵ =1e-8 and (β1, β2) = (0.99, 0.999), as suggested by Ying et al. (2022); Ramp asek et al. (2022). Besides, we also use a learning rate scheduler for graph-level tasks, which is first a linear warm-up stage followed by a cosine decay.