search

xueyunlong12589 / DGCNN

Repetition code of the model for the paper "EEG Emotion Recognition Using Dynamical Graph Convolutional Neural Networks" in pytorch
91 stars 17 forks source link

Why spectral convolution is applied to a directed graph? #5

Open Sunjysama opened 1 year ago

Sunjysama commented 1 year ago

The paper proposed a novel way of learning the weight matrix, however it seems that spectral convolution cannot be applied to a directed graph since the eigenvectors corresponding to the eigenvalues of a non-real symmetric matrix are not necessarily orthogonal.

xueyunlong12589 commented 1 year ago

I am not the author of the paper, and I have my own questions about the place. Here are my own opinions:

From the GCN formula of kipf, it is close to the graph convolution of spatial domain, representing the average aggregation of neighbor nodes. From this perspective, it is reasonable to set the adjacency matrix as a learnable digraph

I have posted the author's home page and public code, you can contact him. You are welcome to continue discussing this issue with me

Sunjysama commented 1 year ago

Many thanks! I have already contacted the author. Let's solve it together.

xueyunlong12589 commented 1 year ago

Ok, please let me know if the author replies.Thanks!

1-2-3-4-0 commented 1 month ago

作者有回复吗