DeosGCF implementation #594
Conversation
|
Btw, the utility functions here https://github.com/RUCAIBox/RecBole/pull/594/files#diff-f5377603dd836d2a11c0891c63600cfe4461a3a3e2deffd88ab05b234bd272daR11-R66 are probably worth extracting out. Some are used e.g. in the LightGCN implementation, but also |
|
Thanks for this implementation. We will review the code and test its performance on our benchmarks. If everything is OK, we will merge it. |
|
Thanks. I assume those are internal benchmarks? If so, I'm curious what the results are |
| A = adjacency_of_bipartite(self.interaction_matrix) | ||
| norm_adj_matrix = get_symm_norm_tensor(A).to(self.device) | ||
| norm_crosshop_matrix = get_symm_norm_tensor(A**2).to(self.device) | ||
| self.A_hat = norm_adj_matrix + norm_crosshop_matrix |
There was a problem hiding this comment.
In the original paper, the matrix is the sum of three matrix: adjacent matrix, cross-hop matrix and identity matrix. It seems that this implementation lack the identity matrix.
| # generate intermediate data | ||
| A = adjacency_of_bipartite(self.interaction_matrix) | ||
| norm_adj_matrix = get_symm_norm_tensor(A).to(self.device) | ||
| norm_crosshop_matrix = get_symm_norm_tensor(A**2).to(self.device) |
There was a problem hiding this comment.
There is a high-pass filtering on the corss-hop metrix which greatly improves the efficiency. It should also be implemented.
3 participants
Paper here: https://arxiv.org/abs/2011.02100
TF implementation here: https://github.com/JimLiu96/DeosciRec
The algorithm is pretty similar to LightGCN so there wasn't too much to do. The paper claims an improvement over LightGCN, which I'm not seeing. So, there might be a mistake somewhere in my implementation.
Named DeosGCF because DGCF is already taken.