search

yao8839836 / text_gcn

Graph Convolutional Networks for Text Classification. AAAI 2019
1.34k stars 434 forks source link

Purpose of computing graph edge weights #23

Open rudra0713 opened 5 years ago

rudra0713 commented 5 years ago

Hi, I have a theoretical question. According to my understanding, the main operations of GCN specified in the paper (Equation 1) are an adjacency matrix, a degree matrix, feature matrix, and weight matrix. How does the edge weights contribute to calculating any of these operations?

yao8839836 commented 5 years ago

@rudra0713

Hi, thanks for the question.

According to [1], GCN is a special form of Laplacian smoothing which computes the new features of a node as the weighted average of itself and its neighbors. Please see equation 8, 9, 10 and their surrounding texts. In equation 9, edge weights are encoded into a˜ij , a larger edge weight between node i and node j will have a larger a˜ij, so when computing the new feature y_i of node i, the feature of node j will contribute more to the calculation.

[1] Li, Q.; Han, Z.; and Wu, X. 2018. Deeper insights into graph convolutional networks for semi-supervised learning. In AAAI.

qinzzz commented 4 years ago

so if i understand correctly, A_ij in the adjacency matrix equals the calculated edge weights in equation (3) in your paper ?

yao8839836 commented 4 years ago

@qinzzz

Yes.