search

yzhang1918 / cikm2021cope

Codes and data for CIKM2021 submission
2 stars 3 forks source link

About the “acgnn.py” file #2

Open jingge326 opened 2 years ago

jingge326 commented 2 years ago

Dear Author,

Thanks very much for sharing this work! I am wondering about the “acgnn.py” file. Would you like to share more details about the ACGNN model? Is there any reference for it? Thanks!

yzhang1918 commented 2 years ago

Hi, jingge326. The ACGNN is the approximate version of CGNN [1]. CGNN is an ODE-based graph neural network, whose closed-form solution is given by the authors (see Eq.7 in [1]). We use Neumann series and Chebyshev polynomials to approximate the closed-from solution. Please refer to Sec 4.4 (Fast Approximation of CGNNs) in our paper [2] for details.

Thanks!

[1] Xhonneux et al., Continuous Graph Neural Networks. ICML'2020. [2] Zhang et al., CoPE: Modeling Continuous Propagation and Evolution on Interaction Graph. CIKM'2021.