Learning Multi-Faceted Prototypical User Interests

Dear Reviewer JT1U,

We would like to thank you for spending time reviewing our paper. We appreciate your constructive comments. We address your concerns as follows

Q1 The proposed method is somehow incremental and is a combination of existing techniques

A1 Regarding the motivation, the proposed method distinguishes itself by two key aspects: discovering multi-faceted item space and modeling the composition of user interests. These motivations have not been considered in [1].

In terms of technical aspects, we designed a novel model including a prototype-based multi-faceted item clustering mechanism, a binding block to model user interest compositionality and the dedicated bi-directional binding block. These components have been empirically shown to be important to achieve higher recommendation accuracy. In contrast, these components were not studied in [1]. Moreover, our model is more general and effective than MacridVAE [1].

As MacridVAE serves as the base of recently disentangled recommendation models, our idea improving MacridVAE, thus, can be applied to a range of problems.

[1] Ma et al. Learning Disentangled Representations for Recommendation. NeurIPS 2019.

Q2 Figure 2 is confusing between explicit user’s rating and implicit user-item interactions

A2 In Figure 2, the input is binary rating vector, where the numbers are IDs of items. Shaded circles are those that user interacted with while dashed circles are unobserved interaction items.

We updated Figure 2 and added clearer notations to ease the understanding.

Q3 Would it be better to include item attribute information in prototype learning as auxiliary information?

A3 It is an interesting direction to incorporate item attributes into the proposed approach. With item attributes such as item categories, one can formulate a classification task to match the learned item cluster with ground-truth categories as classification labels.

In this paper, we consider user-item interaction data for collaborative filtering task, which is the most popular method in recommender systems. As future work, once item attributes are provided, one can integrate them into our proposed approach as aforementioned.

Q4 How does the "latent interest" correspond to explicit interests 1, 2, and 3 in Table 4?

A4 Each user's latent interest, which is denoted by z$^u_k$ (Equation 5) in the main text, is expected to capture one aspect of user preferences. It is derived in an unsupervised manner without any prior information. Thus, to understand the meaning of these latent interests, we find the top 3 items that are closest in representation space with those interests by leveraging output of decoder, i.e., top 3 items with highest decoded score. This score is simply the cosine similarity between latent interest representation and item representations stored in decoder (as shown in Equation 7). Then we visualize the item side information of those items in Table 4 to understand each latent interest.

Dear Reviewer JT1U,

Thank you again for your time reviewing our paper.

We just want to post a friendly reminder that whether our rebuttal has addressed your concerns.

If you have any questions, please let us know. We are happy to discuss.

Best,

Authors.

Dear Reviewer BBhv,

We would like to thank you for your acceptance and your constructive comments. We address your concerns as follows

W1 The proposed approach has worse complexity than MacridVAE

A1 The proposed method outperforms MacridVAE on chosen datasets. Despite having slightly higher complexity than MacridVAE in item clustering step, the complexity is a linear function of the number of items. Moreover, thanks to the multi-faceted item grouping step, FacetVAE can handle large and complex datasets, e.g., with 8 facets and 5 clusters per facet, FacetVAE can describe a dataset including up to $5^8$ (~400k) characteristics. Thus, the proposed method can be applied to large scale datasets.

Q1 In table 1, what is the number of clusters and the number parameters for MacridVAE?

A1 In table 1, we set the number of clusters for MacridVAE is 4, which is also the number of user interests. The number of user interests (K) in FacetVAE is also 4 for fair comparison. In supplementary Figure 4, we varied the number of user interests. In these experiments, FacetVAE still achieves higher recommendation accuracy than MacridVAE.

We configure the settings of MacridVAE and FacetVAE almost the same for fair comparison. The main difference between parameters of these two is multi-faceted clustering step.

FacetVAE uses slightly more parameters than MacridVAE, including $F * J * d^{enc}$ parameters and parameters of $w$ in Equation 3. With small values of $F$ and $J$ (3 or 4) and $d^{enc}$ is not too big (64 or 300), the difference between parameters of MacridVAE and FacetVAE is small. Table below shows the relative comparison w.r.t. the number of parameters between MacridVAE and FacetVAE

Q2 Report the performance with respect to the number of parameters, and computation time across the baselines and the proposed approach

A2 In the answer to Q1 above, we (relatively) compare the number of parameters of FacetVAE and the closest baseline MacridVAE. It is shown that FacetVAE just uses a small number of parameters more than MacridVAE.

In Section 5.2, we already provided the comparison between running time to train single epoch (the reported numbers are averaged over 10 runs) of FacetVAE and MacridVAE. In summary, FacetVAE requires slightly more time than MacridVAE, demonstrating its efficiency.

Q3 In Table 2, why is the number of prototypes fixed to 12 instead of other numbers? Can the best F and J have different trends with different number of prototypes?

A3 Due to limited space, in Table 2, we fix the number of prototypes as 12 for illustration. In supplementary Figure 5, we report the accuracy with various values of number of facets $F$ and number of clusters per facet $J$, which correspond to various values of number of prototypes.

Dear Reviewer BBhv,

May we know that whether your questions have been addressed in the rebuttal above?

Please let us know if you have further questions. We are happy to discuss.

Best Regards,

Authors.

Dear Reviewer Ruv4,

We would like to thank you for spending time reviewing our paper. We appreciate your constructive comments. Your concerns are addressed as follows

W1 How can this method handle and scale to large scale product facets, which can reach up to thousands of facets?

A1 The proposed method can handle large scale datasets including thousands of facets thanks to the multi-faceted modeling. For example, a dataset contains m categories $C_1, C_2 …, C_m$; each includes $F_1, F_2, …, F_m$ facets, respectively. A naïve solution grouping items using single facet would require $F_1 * F_2 * … * F_m$ clusters, which is exponentially large and thus it might be feasible to handle huge datasets. However, the proposed method only requires $F_1 + F_2 + … + F_m$ clusters, which scales linearly. Concretely, if there are 6 categories, each including 4 facets, the total number of required clusters for a naïve solution is $4^6 = 4096$ clusters while the proposed method needs $4 * 6 = 24$ clusters to model item space structure.

W2 This work can be viewed as an incremental upon MacridVAE

A2 Our work is distinct from MacridVAE in three aspects: discovering multi-faceted item space, binding compositional user representations and bi-directional binding block to model compositional user representations. These ideas were not discussed in MacridVAE. Moreover, these aspects have been shown to achieve higher recommendation accuracy.

It is worth noting that MacridVAE serves as the base of recently developed recommendation models. Our key ideas for improving MacridVAE, therefore, are applicable to a wide range of problems.

W3 The performance on large scale (millions of users and items) datasets is unclear

A3 Currently, we are unable to conduct experiments on such huge datasets due to limited computational resources, which we leave these experiments for future work. In any case, the methodology we propose would scale linearly with the number of users or items.

W4 How to leverage product knowledge/other knowledge source to automatically choose #facets should be discussed

A4 When product knowledge, e.g., item category, is available, one could use this information to supervise the item grouping in our proposed method, i.e., a classification task between the learned item group and the ground-truth item category as label.

Dear Reviewer Ruv4,

Given the rebuttal above, may we know that whether we have addressed your concerns about our paper?

We are happy to discuss if you have further questions.

Best Regards,

Authors.