Title: Shattering Information Silos: Leveraging Semantic Personas for Cross-Domain Recommendations

Abstract:

Digital platforms are increasingly functioning as isolated information silos, which restricts their capacity to build holistic user profiles across different domains. While cross-domain recommender systems aim to bypass these limitations by transferring knowledge from a source to a target domain, current methods typically rely on the existence of shared users, common items, or structurally similar interaction graphs. Such prerequisites are frequently unrealistic when dealing with independent platforms. To address this, we introduce SPHERE (Semantic Personas for Heterogeneous cross-domain Recommendation), a framework designed to facilitate knowledge transfer across strictly disjoint domains that share neither users nor items.

Instead of attempting to align domains through identity matching or graph topology, SPHERE employs large language models to establish a common behavioral vocabulary. It generates structured semantic personas for users and retrieves communities from the source domain that exhibit behavioral similarities, creating what is termed a Community Source Persona. This semantic insight is combined with collaborative signals via a dual-tower architecture featuring a dynamic fusion gate, enabling SPHERE to enhance standard recommender system backbones.

Our empirical analysis, conducted across Amazon Books, Goodreads, and Steam datasets, reveals consistent performance gains over baselines such as NCF, SVD++, and LightGCN under full-ranking evaluation conditions. The findings indicate that the efficacy of cross-domain transfer is not dictated exclusively by the semantic closeness of the domains. Instead, it hinges critically on the structural density and inherent predictive power of the target domain. This research contributes to the field of information systems by redefining cross-domain personalization as a process of behavior-based semantic alignment, providing a practical solution for breaking down information silos while maintaining interpretability and modularity.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC