Regarding equation 5 in the paper

[AlwaysFHao]

Hello, regarding the prediction layer formula 5 in the Mamba4Rec paper, $\hat{y} = Softmax(Linear(h)) \in \mathbb{R}^{| \mathcal{V} |}$ I think there is a problem as follows: the calculation of the probability distribution of the prediction layer prediction score in Recbole is not a logic in NLP. It should be based on calculating the similarity (inner product) between the embedding of the final output item and the initial embedding of all selected items to obtain the probability distribution of the prediction score, rather than directly using the linear layer to perform multi classification on the embedding of the final output item.The source code of Mamba4Rec also follows the above rules, which can be found in the calculate'loss and full_sort_predict methods

[AlwaysFHao]

The actual prediction layer formula should be: $\hat{y} = Softmax(h \cdot e) \in \mathbb{R}^{| \mathcal{V} |}$ where e is the original embedding feature of the item corresponding to \hat{y}.

[chengkai-liu]

Thank you for correcting our mistakes. To be accurate, it should be $\hat y=\text{Softmax}\left(h \boldsymbol{E}^\top \right) \in \mathbb{R}^{|\mathcal{V}|}$, where $h \in \mathbb{R}^D$ is the last item embedding from the Mamba layer. $\hat y \in \mathbb{R}^{|\mathcal{V}|}$ represents the probability distribution over the next item in the item set $\mathcal{V}$.

[AlwaysFHao]

Thank you for correcting our mistakes. To be accurate, it should be y^=Softmax(hE⊤)∈R|V|, where h∈RD is the last item embedding from the Mamba layer. y^∈R|V| represents the probability distribution over the next item in the item set V.

Yes, to be precise, the actual formula should be in the form you are giving now. In addition, I have reconstructed Mamba4Rec in a non Recbole environment, but there is still a gap in performance compared to your paper on the Beauty dataset, hit@10 It's 0.0813, but ndcg@10 Only 0.404, could you please run my code and reproduce the effect of Mamba4Rec on Beauty according to your actual parameters? If successful, please let me know?

https://github.com/AlwaysFHao/Mamba4RecWithoutRecbole

[chengkai-liu]

The minor gap in performance is acceptable and within expected variations. It's often challenging to reproduce results exactly across different environments and implementations, even with the same parameters.