Em chào mọi người, hiện em đang nghiên cứu về AP GCN (https://arxiv.org/pdf/2002.10306.pdf) và đang tìm hướng tiếp cận mới tính toán độ tương đồng giữa các node ở phần B, công thức (1), (2), (3), em ko có kiến thức về graph theory nên tìm khó quá, mong các cao nhân chỉ bảo, cảm ơn mọi người ạ!
Firstly, your post is kinda ambiguous. I don’t see your point. Secondly, please forgive me. I literally understand your post but cannot answer in Vietnamese. Sincerely. Hopefully, the below comments are helpful.
My best guess:
- You don’t understand how GCN works? -> In Graph theory, nodes could be represented in an adjacent list. In other words, nodes in a graph could be translated into a binary matrix that 1 represents connection. Watch this https://www.youtube.com/watch?v=5hPfm_uqXmw. In GCN, Convolution is simply applied right on the adjacent matrix (very similar to convolution on image).
- Your concern with equations 1,2,3? -> As explained, GCN is based on Convolution which is usually followed by filters (in CNN, Convolution is usually followed by MaxPool). The equations you mentioned simply describe various filters for GCN and standard GCN formula. In equation 2, the authors describe the Chebyshev filters. Equation 1is the standard weight-updating formula of GCN. Equation 3 simply describes the extension of a general GCN formula.
- How to improve Graph networks? -> My expertise is not in this part yet, and therefore, I cannot suggest you anything on this. Still, please check Graph Attention Networks and this repo: https://github.com/danielegrattarola/spektral. My only suggestion is to find the weakness and drawbacks of AP-GCN and improve it from there. I find it’s hard to improve a general model with goals/objectives.
P/S: the link to the paper is down. Please check .