Hi, I read your amazing work: "Adversarial Graph Disentanglement", and then I have a question about the experimental part. In the V. EXPERIMENTAL RESULTS AND ANALYSIS, E. Analysis of Disentanglement， you compute the component confusion matrix C to illustrate the correlations among the distributions in different component spaces.
The correlation score between the i-th and j-th component distributions is given by
.
I don't understand why this formula is used to calculate the correlation of two vector spaces.
Why is cosine similarity calculated between every two elements in these two spaces？
Why not compute the similarity between corresponding elements as a similarity measure for two vector spaces？that is Cij=1/Vsum(hu^ihu^j).
Finally, can you provide your code on computing the cosine similarity of two vector spaces?
I would appreciate it if you could.
Thank you!