input X should be diagonal matrix equal to A[0]

tkipf / relational-gcn

Keras-based implementation of Relational Graph Convolutional Networks

MIT License

793 stars 134 forks source link

input X should be diagonal matrix equal to A[0] #2

Closed afshinrahimi closed 6 years ago

afshinrahimi commented 6 years ago

Hi,

I enjoyed reading the paper and the code. Good work!

In line 73 of train.py X = sp.csr_matrix(A[0].shape) shouldn't the diagonal be set to 1? I think currently X is a zero matrix.

tkipf commented 6 years ago

Thanks for catching that! In principle, yes, absolutely! In this implementation, however, we use the featureless=True flag for the first layer, which implicitly replaces the input feature matrix with an identity matrix. It's a bit misleading that we still pass an (empty) feature matrix as dummy variable to the model. I will add a comment in the train file.