aikunyi / FourierGNN

Official implementation of the paper "FourierGNN: Rethinking Multivariate Time Series Forecasting from a Pure Graph Perspective"
MIT License
161 stars 15 forks source link

邻接矩阵A的扩散步骤 #8

Open zjucaoyz opened 9 months ago

zjucaoyz commented 9 months ago

您好,我想请问下原文中的邻接矩阵A在代码中是怎么体现迭代的?

Steve-syd commented 9 months ago

同问+1

wangshanhu commented 9 months ago

同问+1, 并没有看见邻接矩阵的输入,这与论文中的FourierGNN(X,A)不相符啊

aikunyi commented 9 months ago

Note that the hypervairate graph is a fully-connected graph.

zjucaoyz commented 9 months ago

Note that the hypervairate graph is a fully-connected graph.

十分感谢回答。A初始化的时候确实都是1,但是第二次迭代是A1(A0XW0)W1,此时A1也都是1吗?论文中似乎提到diffusion step。代码中显示了不同的参数矩阵W,但A似乎一直都是1。

aikunyi commented 9 months ago

因为我们是在hypervariate graph上做的运算,所以A一直是1

Steve-syd commented 9 months ago

因为我们是在hypervariate graph上做的运算,所以A一直是1

请问您这里说的“一直是1”,意思是A一直是单位矩阵还是A一直是全1的邻接矩阵呢?

lqf0624 commented 1 month ago

因为我们是在hypervariate graph上做的运算,所以A一直是1

如果A一直是1,为什么论文里说邻接矩阵是通过FourierGNN学到的?