Aesthetic-Enhanced Multimodal Knowledge Graphs via LLMs for Interoperable Recommender Systems

초록

Beyond text-based representations, multimodal knowledge graphs (MMKGs) integrate visual attributes to represent user preferences and semantic relations among items more effectively. However, with the rapid expansion of multimodal data, achieving interoperability through consistent integration of large-scale heterogeneous graphs is an important aspect to consider for effectively leveraging multimodal information. Nevertheless, modality data are often domain-specific and difficult to connect directly across modalities, making integration within MMKGs difficult to achieve. As a result, most existing studies on interoperability have focused primarily on text-based knowledge graphs. To address these limitations, this study proposes a new framework that enhances interoperability within standardized MMKGs. First, we define AesMMKG, a standardized MMKG that extends the existing standardized text-based schema. Subsequently, we extract three types of aesthetic keywords from visual data using large language models and expand the graph with the proposed algorithm. This extraction process enables seamless alignment between visual features and textual semantic representations. We demonstrate through a case study in real-world recommendation scenarios that the proposed framework enables interoperability across diverse domains. © 2026 Copyright held by the owner/author(s).

키워드