jasperzhong
commented
11 months ago $$X^{(l+1)} = \sigma(D_v^{-1/2}HWD_e^{-1}H^TD_v^{-1/2}X^{(l)}\Theta^{(l)})$$
H是incident matrix ($|V| \times |E|$). 其实Laplacian matrix $L = D - A = H H^T$.
写成message passing的形式是
这里面很大问题是self node feature没有特别考虑，其实self node feature一般才是最重要的.
https://arxiv.org/pdf/1809.09401.pdf