search

tkipf / gcn

Implementation of Graph Convolutional Networks in TensorFlow
MIT License
7.13k stars 2k forks source link

About the data's feature #22

Open songzenghui opened 6 years ago

songzenghui commented 6 years ago

I notice the X in f(X,A) is known.like cora dataset,there's a file named xxx.x or xxx.tx ,how did you get the feature of the cora?If i only have its adj_matrix,how to use your model?

tkipf commented 6 years ago

If you do not have initial features for your nodes, you can either: assign a unique one-hot vector for every node (this results in X being the identity matrix), or you can compute some graph-based features, like node degree, or even a DeepWalk embedding of your nodes. Lastly, you can also combine the two approaches (i.e. concatenate identity matrix with feature matrix) -which can help in some cases.

flyingjohn commented 6 years ago

Hi tkipf, I want to use GCN to model text .So i plan to build a graph with words, and each node represent a word like bird. But in this way , each node represent a dimension and all the nodes consists of the feature of a document. At this time, the channels of each node are the count of doucuments ,does GCN still work?