Unifying User Preferences and Critic Opinions: A Multi-View Cross-Domain Item-sharing Recommender System

We are thankful for your review and the opportunity to clarify aspects of our paper. Here are our responses to your concerns:

1. Sorry for the confusion on data. First we want to clarify that, as detailed in Sections 1 and 4.1, and the Appendix, critic and user reviews in our dataset, sourced from Metacritic, are distinctly labeled, eliminating the need to differentiate between expert and amateur reviews. This dataset will be made publicly available upon paper acceptance.

2. Also we want to clarify that there are no user overlaps in our datasets. As stated in the paper, our work focuses on cross-domain recommendations where different domains share items but not users or critics. The cross-domain in this paper means we leveraging auxiliary information from critic domain data for making better predictions in the user domain. Here 'domain' is defined on the item-level relevance, i.e., there are common items and different users between the targeted and auxiliary domains, as illustrated in the survey paper [1,2]. The auxiliary domain is critic domain and the targeted domain is user domain in this paper. We concur that this specific area remains relatively untapped compared to traditional user-sharing cross-domain recommendation, thus underscoring the significance of our contribution. We will add more explanation in the paper for revision.

The following is our responses to your questions:

Weakness 1. The confusion regarding $diag{\Sigma_{i}}$ is regrettable. In our code, $W_{T}'h_{T}+b_{T}'$ outputs the log variance, which we then convert to positive variance values using the 'torch.exp' function. This is a very frequently used trick in VAE style code. Since this is a common code trick, we did not mention it in the paper. We will add more explanations in the Appendix for revision.

Weakness 2. The activation function in Eq.(1) is tanh in our experiments. We will add more details about it for revision. As introduced in Section 4.1, we set the dropout ratio as 0.5 for the user-rating embedding network.

Weakness 3. We appreciate your correction on the reparameterization formula. This will be amended in the revised paper.

Weakness 4. Thanks for your suggestion. We will incorporate a pseudo algorithm table in the revised paper.

Weakness 5. PTUPCDR assumes that there is overlapping users and no shared items, which is opposite to our focused scenario, i.e., item-sharing and no user overlaps. Therefore, PTUPCDR cannot be directly applied in our problem settings and not included in our baseline methods.

Weakness 6. As mentioned, the distinction between critic and user reviews is inherent in our dataset with labels, negating the need for manual detection.

Weakness 7. In our experiments, there are no overlapping users.

We sincerely appreciate your time and effort in reviewing our work. We hope our responses have satisfactorily addressed your queries. Given the initial misunderstanding regarding the data and cross-domain definition, if you find our clarifications adequate, we kindly request you to consider revising the rating for our paper.

[1] Zang T, Zhu Y, Liu H, et al. A survey on cross-domain recommendation: taxonomies, methods, and future directions[J]. ACM Transactions on Information Systems, 2022, 41(2): 1-39.

[2] Zhu F, Wang Y, Chen C, et al. Cross-domain recommendation: challenges, progress, and prospects[J]. arXiv preprint arXiv:2103.01696, 2021.

Thank you for your response. We have checked them carefully and we have the following responses:

1. Sorry for the confusion on data. First we want to clarify that, as detailed in Sections 1 and 4.1, and the Appendix, critic and user reviews in our dataset, sourced from Metacritic, are distinctly labeled, eliminating the need to differentiate between expert and amateur reviews. This dataset will be made publicly available upon paper acceptance.

2. We apologize for any prior ambiguities regarding the components of the MCIR framework. The MCIR framework is composed of four integral components, each contributing significantly to the methodology:

• User Embedding Networks: This includes two sub-components - the user-rating embedding network and the user-comment embedding network. Their primary function is to learn latent user embeddings from both rating and comment perspectives, respectively.

• Critic Embedding Networks: These networks are designed to concurrently learn latent critic embeddings from both rating and comment perspectives. This dual approach ensures a comprehensive representation of critic opinions.

• Attentive Integrated Mechanism: This mechanism focuses on learning latent item embeddings. It plays a crucial role in transferring auxiliary critic text information into the user domain by enabling shared item embeddings.

• Cross-View Unified Learning with Contrastive Learning: This component is essential for leveraging neighborhood information across user and critic domains. The incorporation of contrastive learning here enhances the effectiveness of using cross-domain neighborhood information to refine the latent user and item embeddings.

The attentive integrated mechanism and cross-view unified learning collectively ensure a robust transfer and utilization of critic information from both the text and graph views. We have provided a comprehensive ablation study in Section 4.3 and the Appendix, which demonstrates the importance of each component. In response to your concerns, we will clarify the main components of the MCIR framework and add more intuitive and descriptive explanations when introducing the modules' details in the revised version of our paper.

3. Regarding Figure 2, the left and right boxes are interconnected via the final latent embeddings $p_i$ and $q_j$, which are trained through both the VAE loss in Equation 9 and the contrastive learning loss in Equation 13. We will try to modify the Figure 2 for better illustration.

4. As outlined in Section 3.3, the latent critic-rating vector $w_l^c$ is derived from the latent critic vector $u_l^c$ using a single-layer MLP network ($w_l^c = f_w(u_l^c)$). The vector $u_l^c$ is initially randomly initialized and then refined through iterative learning.

It appears that there may be some misunderstandings about our work. We sincerely hope that our responses have addressed your concerns and clarified any points of confusion. If you have any further questions or require additional clarification, please do not hesitate to reach out to us.

We are grateful for the opportunity to clarify key points in our paper. Below, we address your concerns and provide further explanations:

1. First, we are sorry for any ambiguity regarding the term 'cross-domain' in our context. Our work indeed focuses on cross-domain recommendations where different domains share items but not users or critics. Critics are not shared by different domains.

Traditionally, cross-domain recommendations have been usually defined in the way that users are sharing across different domains in the literature. However, as evidenced by the survey paper [1] and [2] (please see the Figure 1 and Table 1 in [1]), it's essential to recognize that recommendations integrating distinct user groups with full item overlap also fall under the cross-domain recommendation umbrella [1][2][3]. As stated in the paper, the cross-domain in this paper means we leveraging auxiliary information from critic domain data for making better predictions in the user domain. Here 'domain' is defined on the item-level relevance, i.e., there are common items and different users between the targeted and auxiliary domains, as illustrated in the survey paper [2]. The auxiliary domain is critic domain and the targeted domain is user domain in this paper. We concur that this specific area remains relatively untapped compared to traditional user-sharing cross-domain recommendation, thus underscoring the significance of our contribution. We will add more explanation in the paper for revision.

2. In our experiments, we train and test the model on each dataset independently. There is no data sharing among the three datasets. MCIR can outperform all the baseline methods, demonstrating the effectiveness of our cross domain leaning approach.

The following is our responses to your questions:

Question 1. Thank you. $y_{lj}$ should be $y_{lj}^c$ in the left up corner of Figure 2. We will correct this in the revised paper.

Question 2. As stated in the paper, the user embedding network belongs to the encoder part of the VAE architecture. Hence we use $R_i$ to construct $u_i$ but not using $u_i$ to generate $R_i$. The latter is suitable for the decoder part of VAE. We will add more explanation in the revised paper.

Question 3. Sorry for the confusion around the introduction of $L_j$ and $w_{0}^c=v_j$ in Equation 6. These elements are meant to simplify the mathematical expression. Without the help of $L_j$ and $w_{0}^c=v_j$, Equation 6 will look much more complex. According to your comment, we will refine our explanation of the attentive integration mechanism, particularly focusing on the roles of key, query, and value vectors.

Question 4. As previously mentioned, our study defines 'cross-domain' as leveraging critic domain data to enhance user domain predictions. Critics, often from professional media institutions, and users, typically ordinary web contributors, represent distinct groups without overlap. Our study bridges these 'user' and 'critic' domains in a novel cross-domain approach.

We sincerely appreciate your time and effort in reviewing our work. We hope our responses have satisfactorily addressed your queries. Given the initial misunderstanding regarding the cross-domain definition, if you find our clarifications adequate, we kindly request you to consider revising the rating for our paper.

[1] Zang T, Zhu Y, Liu H, et al. A survey on cross-domain recommendation: taxonomies, methods, and future directions[J]. ACM Transactions on Information Systems, 2022, 41(2): 1-39.

[2] Zhu F, Wang Y, Chen C, et al. Cross-domain recommendation: challenges, progress, and prospects[J]. arXiv preprint arXiv:2103.01696, 2021.

[3] Gao C, Li Y, Feng F, et al. Cross-domain recommendation with bridge-item embeddings[J]. ACM Transactions on Knowledge Discovery from Data (TKDD), 2021, 16(1): 1-23.

We greatly appreciate your insightful feedback and recognition of the strengths in our paper, particularly the well-motivated problem statement and our system's ability to surpass state-of-the-art results. Below, we have addressed your concerns in detail:

Weakness 1: The attentive integration mechanism is essential in combining item-summary information with critics' commentary as it adaptively aggregates information, reflecting the varying influences of item summaries and critics’ comments on item characteristics. We have included two case studies in Appendix Section D, demonstrating its effectiveness. In response to your suggestion, we introduce a new variant, MCIR-C6, employing simple average of item-summary vector $v_j^s$ and critic vectors $w_{l}^c$ instead of attentive integration. The comparative results on MC-Game are as follows:

Methods R@5 P@5 MAP@5 NDCG@5

MCIR-C6 0.1050 0.0845 0.0545 0.1172

These results clearly show the superior performance of MCIR over MCIR-C6, affirming the value of the attentive integration mechanism. Regarding the query “what if all $v_j^a=0$?”, in practice, $v_j^a$ will not be zero since the item-summary vector is always present. The variant MCIR-C2, which replaces $v_j^a$ with the item-summary vector $v_j$, underperforms compared to MCIR as shown in Figure 3(a). This further highlights the importance of the review text information alongside the attentive integration mechanism. We plan to include MCIR-C6 in our revised ablation study and provide an in-depth discussion of the attentive integration mechanism.

Weakness 2: Actually, the variant MCIR-C4 is set as $\eta_3=\eta_4=0$. Note that the graph embeddings can only influence the final latent user and item representation $p_i$ and $q_j$ by the contrastive learning loss $\mathcal{L}_{CL}$.

So when we set $\eta_4=0$, the GCN loss $\mathcal{L}_{Graph}$ is irrelevant to the predictions, which can also be viewed as $\eta_3=0$. We will elaborate on this in the revised manuscript.

Weakness 3: In our study, we focus on leveraging the auxiliary critic domain to enhance user domain predictions. To avoid overcomplicating the probabilistic derivation process, we did not use a VAE in the critic embedding network. Note that $v_j^a$ is involved in the loss function $\mathcal{L}_{VAE}$, implicating $u_l^c$ in the VAE's ELBO for $u_i$. Employing two VAEs for $u_i$ and $u_l^c$ would significantly complicate the VAE Loss due to their interdependence. Thus, we opted for a VAE solely in the user domain but not the auxiliary critic domain.

As for your concern “why not use the (analogous) first term in equation of 5 in equation 3”, this is because equation 5 is for concurrently learning the latent critic-rating and critic-comment vectors while equation 3 is only for learning the User-Comment vectors. The first term in equation of 5 is for learning critic-rating vectors. As discussed in the paper, given the diverse nature of user comments and their potential limited correlation with item property evaluations [1,2,3] and significant correlation exists between critics' scores and their explanatory comments [1,2], we opt for independent learning for users while concurrent learning for critics. According to your comments, we will expand on the rationale behind the loss design in our revised version.

Response to Questions:

Question 1: We apologize for the typos. $\eta_s$ should be $\eta_1$, $\eta_c$ should be $\eta_2$, and $\eta_n$ should be $\eta_4$. These will be corrected in our revised paper.

Question 2: Sorry for the confusion on the ablation study C1. It excludes review text information for both user and critic embedding networks. Specifically, we set the user-comment vector $w_i=0$, and the loss $\mathcal{L}_{MSE}$ is not applied.

Additionally, the term $| y_{lj}^c-z_{lj}^c| ^2$ is omitted in the loss $\mathcal{L}_{Multi}$. We also replace $v_j^a$ with the item-summary vector $v_j$. More details will be provided in the revised version.

Question 3: Ablation study C2 involves replacing $v_j^a$ with the item-summary vector $v_j$.

Question 4: Ablation study C3 involves not applying the loss $\mathcal{L}_{Multi}$ and replacing $v_j^a$ with $v_j$.

Additionally, all critic-item edges are omitted from the graph for the loss $\mathcal{L}_{Graph}$.

Question 5: Yes. Ablation study C4 mean $\eta_4=0$.

Owing to the constraints of space, we will provide the remaining responses in our subsequent reply.