Request for KAR Code

[Smingaeng]

Hello,

Thank you for your great work on this paper.
I noticed that in your experiments you compared against KAR, originally proposed as a reranker model. Could you kindly let us know if you plan to release the KAR code you used for comparison?

In particular, I am interested in the implementation details, since the original evaluation setting of KAR seems different from the one used in your paper. Could you clarify how you handled this in your experiments? Any implementation details or code release would be very helpful for reproducing your results.

Thank you very much for your time and consideration.

[Re-bin]

Hi 👋!

Thank you for your interest in RLMRec. We are pleased to provide the following implementation details:

• We implemented KAR within the collaborative filtering (CF) scenario by adopting its core ideas of knowledge adaptation and knowledge utilization.
• Knowledge adaptation: We use the same semantic embeddings for users and items as in RLMRec. Then, the preference/share experts and gating networks are implemented as MLPs to process these embeddings, resulting in augmented semantic embeddings for each user and item.
• Knowledge utilization: As in the original paper, the augmented vectors serve as inputs to the downstream model. Concretely, we concatenate each augmented semantic embedding with its corresponding CF embedding, and then pass the concatenated vectors to the downstream CF model to obtain final user/item embeddings for prediction, consistent with other CF models.

We hope this information is helpful to you:)

Best regards,
Xubin