A Method to Manage the Precision Difference between Items and Profiles: In a Context of Content-Based Recommender System and Vector Space Model

Contractors, commercial and business decision-makers need economical information to drive their decisions. The production and distribution of a press review about French regional economic actors represents a tool for prospecting partners and competitors for the businessman. Our goal is to propose a customized review for each user, thus reducing the overload of useless information. Several news recommendation systems already exist. The usefulness of external knowledge sources to improve the process has already been explained in information retrieval. The system's knowledge base includes the domain knowledge used during the recommendation process. Our recommender system's architecture is standard, but during the indexing task, the representations for each article's content and for each interest from the users' profiles created are based on this domain knowledge. Articles and Profiles are semantically defined in the Knowledge base via concepts, instances and relations. This paper deals with the relevance measure, a critical sub-task in recommendation systems and the relationships between this metric and similarity concepts. The Vector Space Model is a well-known model used for relevance ranking. First, the problem exposed here concerns the use of the standard VSM method along with our indexing method. Then, we test our approach by means of practical tests, such as classic precision recall, and more specific evaluation methods to compare ranking results. Finally, we expose evaluation results of our recommendation algorithm and discuss the case of multilabeled items.