why need transpose in adjacency matrix normalization?

[yanlirock]

in the function normalize_adj the output is: "adj.dot(d_mat_inv_sqrt).transpose().dot(d_mat_inv_sqrt).tocoo()"

why not using "d_mat_inv_sqrt.dot(adj).dot(d_mat_inv_sqrt) " follow by the original equation in paper?

and the results of above two are different

[MeetUo]

I also have this confusion. But I find the adj matrix is the symmetric matrix and the d_mat_inv_sqrt is the diagonal matrix . So "adj.dot(d_mat_inv_sqrt).transpose().dot(d_mat_inv_sqrt)" actually equal "d_mat_inv_sqrt.dot(adj).dot(d_mat_inv_sqrt) " But I also don't know why need transpose, hope the author can answer our doubts.

[leeswan]

transpose for construct symmetric normalized adj matrix in tensorflow version, while in pytorch version, the method of normalization is different, features and adj take the same way

[kentfly]

(A*B)T = BT * AT so.. "adj.dot(d_mat_inv_sqrt).transpose().dot(d_mat_inv_sqrt)" actually not equal to "d_mat_inv_sqrt.dot(adj).dot(d_mat_inv_sqrt) "

I hope the author can explain this problem which is so important for the GCN...thx...

I also have this confusion. But I find the adj matrix is the symmetric matrix and the d_mat_inv_sqrt is the diagonal matrix . So "adj.dot(d_mat_inv_sqrt).transpose().dot(d_mat_inv_sqrt)" actually equal "d_mat_inv_sqrt.dot(adj).dot(d_mat_inv_sqrt) " But I also don't know why need transpose, hope the author can answer our doubts.

[tkipf]

If adj is symmetric (which we assume) then both are identical.

On Thu, Jan 2, 2020 at 10:30 AM kentfly [email protected] wrote:

(A*B)T = BT * AT so.. "adj.dot(d_mat_inv_sqrt).transpose().dot(d_mat_inv_sqrt)" actually not equal to "d_mat_inv_sqrt.dot(adj).dot(d_mat_inv_sqrt) "

I hope the author can explain this problem which is so important for the GCN...thx...

I also have this confusion. But I find the adj matrix is the symmetric matrix and the d_mat_inv_sqrt is the diagonal matrix . So "adj.dot(d_mat_inv_sqrt).transpose().dot(d_mat_inv_sqrt)" actually equal "d_mat_inv_sqrt.dot(adj).dot(d_mat_inv_sqrt) " But I also don't know why need transpose, hope the author can answer our doubts.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/tkipf/gcn/issues/142?email_source=notifications&email_token=ABYBYYA2J4NMERD46Z46ZYTQ3WX23A5CNFSM4I25WHO2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEH56BLQ#issuecomment-570155182, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABYBYYC4IUX5DG3N6EUKRNLQ3WX23ANCNFSM4I25WHOQ .

[kentfly]

i get it!! thx a looooot....:))