What is gr_output_length specifically used for?

[yudonglee]

Hi, I am very confused about the specific role of gr_output_length.

The length of hstu's input seq is (max_sequence_length + gr_output_length + 1),
For the 'gr_output_length' part, the corresponding historical_id is 0, as shown in the code below：

def movielens_seq_features_from_row(
    row: Dict[str, torch.Tensor],
    device: int,
    max_output_length: int,
) -> Tuple[SequentialFeatures, torch.Tensor, torch.Tensor]:
... ...
        B = historical_lengths.size(0)
        historical_ids = torch.cat(
            [
                historical_ids,
                torch.zeros(
                    (B, max_output_length), dtype=historical_ids.dtype, device=device
                ),
            ],
            dim=1,
        )
... ...

Since the corresponding historical_ids is 0, this part seems to have no effect on the final calculation result. So what is its purpose?
I wonder if it is related to your paper, where gr_output_length corresponds to the number of vectors in the user's multi-vector representation?

https://arxiv.org/pdf/2306.04039

[jiaqizhai]

Hi, the 0 padding in the id sequence is due to our specific implementation of relative position bias. We use timestamp[j+1] - timestamp[i] to derive the delta time span from i to j (note here we autoregressively predict the j+1-th element) in our public experiments for sequential recommender (retrieval) settings.