| （英） |
Current personalized information services offered by large platform providers typically keep content, user data, recommendation policies, and the interpretation of recommendation results within a single closed service, making it difficult for users to select or adjust how information is explored. In this research, we propose a user-sovereign, open information infrastructure that separates an information service into four components: content providers, knowledge providers, explorers, and explainers. By standardizing the input and output interfaces of these components, providers become interchangeable, allowing users to freely combine components from different parties according to their own purposes. In particular, content providers supply the set of items to be explored and their identifiers. Knowledge providers use their internal knowledge to provide feature vectors representing the characteristics of each item. Explorers combine multiple feature vectors to generate a final recommendation vector. Explainers use large language models to describe the recommendation results and the process by which they were controlled in natural language. As a case study of the proposed framework, we present a movie recommendation prototype that integrates vectors representing popularity, genre, tag relevance, and proximity derived from a rating dataset and the link structure of Wikipedia, and then uses an LLM to interpret the resulting recommendations. |