Can I think graph convolution as 2D convolution like images?


Kipf et al described in his paper that we can write graph convolution operation like this:

$$H_{t+1} = AH_tW_t$$

where, $A$ is the normalized adjacency matrix, $H_t$ is the embedded representation of the nodes and $W_t$ is the weight matrix.

Now, can I imagine the same formula as first performing 2D convolution with fixed-size kernel over the whole feature space then multiply the result with the adjacency matrix? If this is the case, I think I can create graph convolution operation just using the Conv2D layer then performing simple matrix multiplication with adjacency matrix using PyTorch.

Swakshar Deb

Posted 2020-07-24T08:58:11.450

Reputation: 432


– nbro – 2020-07-24T19:49:30.033

No answers