ScholarWorks@숙명여자대학교

초록

Recommender systems are now ubiquitous in e-commerce and review platforms, yet they often struggle when interaction data is extremely sparse (the so-called cold-start problem). To alleviate this, many approaches incorporate auxiliary side information (e.g., social links or item attributes) and exploit transfer learning from related domains. In particular, user-provided reviews have been widely used to enrich sparse feedback data. However, existing review-based cross-domain methods typically assume a Euclidean latent space, which can poorly capture the inherent hierarchical, power-law nature of user-item interactions. In Euclidean space, gradient updates typically affect distances in a roughly linear fashion. In contrast, hyperbolic distances grow approximately exponentially with radius, so naïvely aligning embeddings in tangent space can drastically change the relative radial positions of nodes. This may push popular and rare entities too close together near the origin and collapse the intended hierarchy. In this work, we propose a novel framework for review-based cross-domain recommendation that operates in hyperbolic geometry and explicitly preserves hierarchical structure. Our model (HEAD) embeds users and items into a hyperbolic manifold and uses carefully designed alignment procedures: a degree-based normalization aligns node distances to emphasize popular (high-degree) entities, and a scale-normalized domain-discriminator ensures stable cross-domain feature matching (see Theoretical Analysis). These hierarchy-aware schemes avoid the collapse of the exponential hyperbolic volume under naive alignment. We validate our approach with both theoretical insights and extensive experiments, showing that HEAD achieves superior accuracy and robustness compared to Euclidean and unaligned baselines.

키워드