目录
Graph Signal Processing and Spectral-based GNN
-
Introduction
Graph是由节点和边组成的,节点有节点的性质,边有边的性质:
Graph Neural Networks可以做的事情:Classification、Generation。
How to embed node into a feature space using convolution?
- Solution 1: Generalize the concept of convolution (corelation) tograph >> Spatial-based convolution
- Solution 2: Back to the definition of convolution in signal processing>> Spectral-based convolution。
-
Spatial-based GNN
下面介绍下实际大部分的model是怎么做的:
NN4G (Neural Networks for Graph)
首先做Aggregate:
然后做readout:
DCNN (Diffusion-Convolution Neural Network)
首先做Aggregate:
然后做readout:
DGC (Diffusion Graph Convolution)
Aggregate的步骤和DCNN一样,readout的做法不同。
MoNET (Mixture Model Networks)
GAT (Graph Attention Networks): GAT要做weighted sum,并且weighted sum的weight要通过学习得到。
GIN (GraphIsomorphism Network)
GIN在理论上证明了有些方法会work,有些方法不会work。
GIN的结论告诉我们,在update的时候,最好可以用下面的方式做update:
-
Graph Signal Processing and Spectral-based GNN
Spectral-Based Convolution:
Spectral Convolution的一些model:
① ChebNet 速度很快而且可以localize,但是它要解决time complexity太高昂的问题。
它通过用Chebyshev polynomial来解决time complexity太高昂这个问题。
② GCN