Disentangled Heterogeneous Collaborative Filtering

Responses to Reviewer KKy5

Thank you for your dedicated efforts and valuable feedback. We would like to address your concerns as follows:

1. The practicality of the proposed DHCF model (W1)

• Model Efficiency. In our efficiency study, as presented in Section 4.6 and Table 2, we demonstrate that the per-epoch training and testing time of DHCF is comparable to existing baseline models. This indicates that DHCF is applicable in real-world scenarios and is efficient in terms of computational resources.

• Model Convergence. Regarding the convergence of the model, our DHCF follows the same maximum training epochs as other baseline methods, which is set to 100 epochs. This suggests that there are no disadvantages in terms of convergence speed for DHCF compared to other models.

2. Further clarification of the NCF's results on different user groups (in Figure 3). (W2 & Q1)

   Thank you for your insightful feedback. We appreciate your comment and would like to address it with the following points:

• Increment of supervision signals does not always improve performance. The observation that basic methods like NCF achieve similar performance on different user groups, as shown in Figure 3, is consistent with findings from previous works on multi-behavior recommendation, such as [1-2]. This phenomenon highlights that simply increasing supervision signals does not necessarily lead to improved model performance. The lack of high-quality self-supervision signals can limit the effectiveness of increasing supervision, and this issue can also affect self-supervised learning methods based on random augmentations.

• Consistent performance gain of DHCF. In contrast to existing supervision-enhancing methods like SGL, our DHCF achieves performance improvement across different data sparsity levels through intent disentanglement and modeling of heterogeneous relation dependencies. This design effectively enhances the overall model performance. The consistent performance superiority of DHCF over the baseline models further validates the effectiveness of these designs.

  [1] "Knowledge Enhancement for Contrastive Multi-Behavior Recommendation"

  [2] "Graph Meta Network for Multi-Behavior Recommendation"

3. Performance evaluation on tailed items (Q1)

• Thank you for your suggestion. We appreciate your interest in evaluating our model's performance on tailed items. In our experiment, we specifically considered a user set composed of long-tailed users to address the issue of sparsity. These users have a maximum of 35 interactions across different types, with an average of only 3.81 purchase interactions. This indicates that these users have a significantly low number of interactions. While we did not explicitly mention "tailed items" in our paper, our analysis does encompass the evaluation of our model's performance on this particular long-tailed user set.

4. Is the meta-learning process considered while calculating the complexity in section 3.7? (Q2)

• We appreciate your thoughtful question. Although we did not explicitly mention the complexity of the meta network in our paper for the sake of simplicity, it involves a straightforward linear transformation. This transformation incurs a cost of $O((I+J)\times d^2)$ for each of the $K$ heterogeneous relations. Empirically, we have found that this complexity is close to $O(|\mathcal{X}|\times d)$, which is smaller than the complexity of the graph relation learning process. The graph relation learning process has a cost of $O(|\mathcal{X}|\times d\times L)$, where $|\mathcal{X}|$ represents the size of the dataset and $L$ denotes the number of graph convolutional layers.

5. Discussion of model convergence with the learnable hypergraphs? (Q3)

• Thank you for your question. During the training process of DHCF, we have noticed that the performance on the validation set remains stable in the last few epochs. This consistent behavior can be replicated using the code we have released, providing strong evidence for the convergence of the learnable hypergraphs.

Responses to Reviewer 7WGw

Thank you for your kind feedback. We appreciate your concerns, and we would like to address them as follows:

1. Further clarifications of difference between the new model over existing methods. (W1)

• We appreciate your feedback. In our work, we address two important challenges that have been overlooked in the field of collaborative filtering: interaction heterogeneity and fine-grained modeling of user intents. To bridge these gaps, we propose DHCF, a method that integrates intent disentanglement and multi-behavior modeling using a parameterized heterogeneous hypergraph architecture. Furthermore, we introduce a novel approach to heterogeneous contrastive learning, leveraging the encoded disentangled heterogeneous interaction patterns.

2. Further discussion of some details in the methodology part. (W2)

• Thank you for your feedback. To better address your concern, we kindly request specific feedback on the parts that require further clarification. This will enable us to identify and improve those specific areas more effectively. Please let us know if there are any specific questions or sections where you would like us to provide more details.

3. Clarifications about the already conducted scalability and hyperparameter studies. (W3)

• We appreciate your thoughtful suggestion. In our work, we have thoroughly investigated the influence of hyperparameter settings in Section 4.5 and examined model scalability in Section 4.6. Additionally, we have conducted comprehensive experiments covering various aspects. Specifically, we have performed a rigorous performance comparison with 18 diverse baselines, explored the impact of hyperparameter settings, conducted module ablation studies, assessed model scalability, evaluated robustness against data sparsity, and presented a detailed case study on the learned hypergraph structures.

4. How does DHCF exploit the relational dependencies across different types of user behaviors? (Q1)

• Thank you for providing the feedback. In DHCF, we integrate cross-type contrastive learning tasks at both the node level and the graph level to leverage the relational dependencies across different types of user behaviors. At the node level, our contrastive learning approach aligns the learned node representations across diverse heterogeneous interaction types, facilitating the identification of shared characteristics across different behavior modes. Additionally, at the graph level, our contrastive learning aligns the representations for the entire graph, taking into account the cross-type commonalities in global graph structures.

5. How does the parameterized heterogeneous hypergraph architecture capture the complex and diverse interactions among users, items, and behaviors? (Q1)

• Our parameterized hypergraph networks capture user-wise and item-wise relations from a global perspective by utilizing hypergraph-based global relation learning for each behavior type. This approach enables us to effectively model the heterogeneous interactions and fuse multi-type representations, thereby facilitating the modeling of relational dependencies between different entities.

6. How does DHCF disentangle the latent intent factors behind each behavior? (Q2)

• DHCF leverages a parameterized hypergraph neural network to disentangle the latent intent factors that underlie each user behavior. In this architecture, the hyperedges in the hypergraph represent the latent user intents. Through iterative hypergraph structure learning, DHCF establishes connections between users/items and their associated hyperedges, effectively capturing users' frequent intents. This disentanglement process enables DHCF to model and differentiate the underlying intent factors that drive different user behaviors.

7. How does the behavior-wise contrastive learning paradigm facilitate adaptive data augmentation at both the node and graph levels? (Q2)

• In DHCF, the adaptive data augmentation primarily targets the node level and involves personalized transformations tailored to different users and behavior types. Specifically, DHCF incorporates user-specific embedding transformations for their type-specific embeddings, which enables accommodating user-specific variations in behavior types. This approach allows DHCF to adaptively augment the data at a personalized level, enhancing the modeling of user behaviors with consideration for individual differences.

Responses to Reviewer TWFu

Thank you for your valuable feedback. We would like to address your comments as follows:

1. Further discussion about the our encoder novlty. (W1 & Q)

Our methodology comprises two primary components: the encoding part, which utilizes a heterogeneous hypergraph neural network, and the self-supervised learning part, employing heterogeneous contrastive learning. In order to address these concerns and highlight the technical novelty of our DHCF model, we provide the following justifications:

• Uniqueness of DHCF's Encoding Part. While our DHCF model and existing works such as HCCF share some commonalities (e.g., the use of hypergraph neural networks), there are two significant technical differences in the design of our encoder: heterogeneity and disentanglement. Unlike existing hypergraph-based methods like HCCF, which primarily focus on homogeneous data, our DHCF model addresses the challenge of heterogeneity by considering interaction types and effectively preserving heterogeneity information in the hypergraph-based encoder. Furthermore, the adoption of a parameterized hypergraph neural network in our DHCF model facilitates efficient and learnable intent disentanglement. This disentanglement is vital not only during the encoding process but also in the subsequent self-supervised learning part, as discussed below. These technical distinctions highlight the unique contributions of our DHCF model, enabling it to handle heterogeneity and disentangle user intents more effectively compared to existing approaches.

• Novelty of Our Contrastive Learning and Its Relation to Encoder. Our paper introduces a significant technical contribution in the form of the proposed disentangled heterogeneous contrastive learning method. This method incorporates several key components, including a parameter personalization module, a node-level heterogeneous self-discrimination task, and a graph-level contrastive learning task. These components leverage hypergraph-based global representation and shuffling-based negative sample construction techniques. By employing these designed techniques, our method greatly enhances the training of models for heterogeneous collaborative filtering. It achieves this by facilitating adaptive supervision enhancement while considering global intent disentanglement. Notably, our approach outperforms vanilla graph contrastive learning methods such as SGL, primarily due to the valuable heterogeneity and disentanglement information extracted by our designed encoding part.

• Theoretical Contribution. Our paper makes a contribution by providing an in-depth theoretical analysis that highlights two important advantages of DHCF that distinguish it from existing works: i) hypergraph-based disentanglement enhances the adaptability of contrastive supervision signals, ii) graph-level contrastive learning injects global disentangled representations into node-level similarities. Thank you for your insightful feedback. We will further highlight the novelty and uniqueness of our encoding component in future versions.

2. Further theoretical justification of SSL loss terms (W2)

• The two SSL loss functions employed in our study follow the contrastive learning paradigm. However, they differ in their specific focuses and objectives. The first SSL loss function concentrates on micro-scale learning, targeting the individual nodes within the graph. It enables fine-grained learning at the node level, capturing local heterogenous collaborative relationships among users and items. The second SSL loss function places emphasis on macro-scale learning, aiming to maximize learning for the overall graph structure. This macro-scale approach ensures that the model captures the global context and structural information within the multi-behavior data. Furthermore, since natural negative samples are lacking at the graph level, we introduce a shuffling-based negative sampling generation method, which effectively enlarges the technical similarity between the two loss functions.

3. Font size of Fig.1 is too small. (W3)

• Thank you for your thoughtful suggestion. We will improve the clarity by updating the figure in later versions.