why would you update item_embeddings with the linear transformation of o.

[huanzhao595]

`self.transform_matrix = nn.Linear(self.dim, self.dim, bias=False)

item_embeddings = self._update_item_embedding(item_embeddings, o)

item_embeddings = self.transform_matrix(item_embeddings + o)
`
As the codes above show, vector o can be seen as the 1-order response of user u’s click history Vu with respect to the candidate item. Item_embeddings can be seen as the candidate item embeddings.

I have a question, why would you update item_embeddings with the linear transformation of o.

To be honest, I drop this code and the precision decreased dramatically, the highest precision is close to 50% on Movielens. So what's the purpose of this linear transformation.

PS, I didn't find any theory in the paper to explain this code. Please tell me why.

[zhoupengfeigjqh]

+1