Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Review-Based Hyperbolic Cross-Domain Recommendation

Yoonhyuk Choi, Jiho Choi, Taewook Ko, Chong-Kwon Kim

TL;DR

The paper tackles data sparsity in recommender systems by integrating review-based signals with cross-domain knowledge transfer in a hyperbolic space. It introduces HEAD, a Hyperbolic Embedding and Hierarchy-Aware Domain Disentanglement framework that uses degree-based normalization and scale alignment to preserve hierarchical structure while separating domain-specific and domain-shareable information. The approach combines Poincaré Glove text representations, multi-channel CNN feature extractors, and hyperbolic distance-based prediction with a margin ranking objective, achieving strong performance on 12 Amazon domain pairs and demonstrating improved domain transfer robustness. This work establishes hyperbolic geometry as a principled foundation for cross-domain, review-informed recommendations, offering scalable and interpretable hierarchy-aware knowledge transfer for sparse settings.

Abstract

The issue of data sparsity poses a significant challenge to recommender systems. In response to this, algorithms that leverage side information such as review texts have been proposed. Furthermore, Cross-Domain Recommendation (CDR), which captures domain-shareable knowledge and transfers it from a richer domain (source) to a sparser one (target), has received notable attention. Nevertheless, the majority of existing methodologies assume a Euclidean embedding space, encountering difficulties in accurately representing richer text information and managing complex interactions between users and items. This paper advocates a hyperbolic CDR approach based on review texts for modeling user-item relationships. We first emphasize that conventional distance-based domain alignment techniques may cause problems because small modifications in hyperbolic geometry result in magnified perturbations, ultimately leading to the collapse of hierarchical structures. To address this challenge, we propose hierarchy-aware embedding and domain alignment schemes that adjust the scale to extract domain-shareable information without disrupting structural forms. The process involves the initial embedding of review texts in hyperbolic space, followed by feature extraction incorporating degree-based normalization and structure alignment. We conducted extensive experiments to substantiate the efficiency, robustness, and scalability of our proposed model in comparison to state-of-the-art baselines.

Review-Based Hyperbolic Cross-Domain Recommendation

TL;DR

The paper tackles data sparsity in recommender systems by integrating review-based signals with cross-domain knowledge transfer in a hyperbolic space. It introduces HEAD, a Hyperbolic Embedding and Hierarchy-Aware Domain Disentanglement framework that uses degree-based normalization and scale alignment to preserve hierarchical structure while separating domain-specific and domain-shareable information. The approach combines Poincaré Glove text representations, multi-channel CNN feature extractors, and hyperbolic distance-based prediction with a margin ranking objective, achieving strong performance on 12 Amazon domain pairs and demonstrating improved domain transfer robustness. This work establishes hyperbolic geometry as a principled foundation for cross-domain, review-informed recommendations, offering scalable and interpretable hierarchy-aware knowledge transfer for sparse settings.

Abstract

The issue of data sparsity poses a significant challenge to recommender systems. In response to this, algorithms that leverage side information such as review texts have been proposed. Furthermore, Cross-Domain Recommendation (CDR), which captures domain-shareable knowledge and transfers it from a richer domain (source) to a sparser one (target), has received notable attention. Nevertheless, the majority of existing methodologies assume a Euclidean embedding space, encountering difficulties in accurately representing richer text information and managing complex interactions between users and items. This paper advocates a hyperbolic CDR approach based on review texts for modeling user-item relationships. We first emphasize that conventional distance-based domain alignment techniques may cause problems because small modifications in hyperbolic geometry result in magnified perturbations, ultimately leading to the collapse of hierarchical structures. To address this challenge, we propose hierarchy-aware embedding and domain alignment schemes that adjust the scale to extract domain-shareable information without disrupting structural forms. The process involves the initial embedding of review texts in hyperbolic space, followed by feature extraction incorporating degree-based normalization and structure alignment. We conducted extensive experiments to substantiate the efficiency, robustness, and scalability of our proposed model in comparison to state-of-the-art baselines.
Paper Structure (23 sections, 25 equations, 5 figures, 2 tables)

This paper contains 23 sections, 25 equations, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Review-Based Recommendation
  4. Cross-Domain Recommendation
  5. Hyperbolic Recommendation
  6. Preliminaries
  7. Methodology
  8. Word Embedding
  9. Feature Extraction
  10. Hierarchy-Aware Hyperbolic Embedding and Domain Disentanglement
  11. Hierarchy-aware hyperbolic embedding.
  12. Hierarchy-aware domain disentanglement.
  13. Inference and Optimization
  14. Time Complexity
  15. Theoretical Analysis
  16. ...and 8 more sections

Figures (5)

  • Figure 1: The geometric properties of hyperbolic space require that popular (or most interacted) nodes be placed near the origin. Let us assume that user $u_1$ purchased an item $i_1$ (left). Given that general algorithms bring relevant nodes closer together, after the update, a structural collapse occurs (right) as $i_1$ moves farther from the origin
  • Figure 2: The overall framework of the Hierarchy-Aware Hyperbolic Embedding and Domain Disentanglement (HEAD) scheme. The (1)-(3) represents three types of loss functions
  • Figure 3: (RQ2) Domain discrimination performance of three methods with similar and dissimilar domain pairs
  • Figure 4: (RQ3) We randomly sampled 1,000 items in Digital Music and visualized them based on their degrees
  • Figure 5: (RQ4) We describe the NDCG@10 of two datasets by varying the parameters $\lambda_1$ (x-axis, degree normalization) and $\lambda_2$ (y-axis, scale alignment) in Eq. \ref{['total_loss']}, respectively