Uma abordagem baseada em linked open data para diversificação de recomendações.

Abstract

Diversity is an important concept in Recommender Systems (RS) since diversified recommendations can help the users to find more interesting and relevant items. Diversity in RS is mainly achieved considering two aspects: (i) the similarity between the items in the recommendations list, under the assumption that the more dissimilar, more diverse is the list; and (ii) the coverage of items’ attributes, i.e., the more attributes covered in the recommendation list (e.g., musical genres on musical artists recommendation), the more diverse is the list. Given that it is not always easy to access or extract the attributes of the items of a RS, (i) is still the predominant approach exploited in the literature. However, thanks to the semantic Web and Linked Open Data (LOD) initiatives, several attributes commonly found in many recommendation domains (e.g. movies, books and music) are now publicly available in RDF databases, connected to each other in the LOD Cloud. In this work, we propose and implement a new approach for promoting diversification in RS that exploits the semantic relationships between item’s attributes, both extracted from LOD repositories, found in LOD repositories as well as it several content dimensions. Another contribution of this work is that we tackle the typical trade-off between accurate and diversified recommendations by inferring the degree of diversification (used as a diversifier parameter) directly from the user profile. We conduct a thorough evaluation of our approach on real data collected from Last.FM, a social, online radio station and scrobbler, which stores online history of userheard songs. It has been shown that the proposed approach complements and exceeds the work related to diversity and accuracy metrics, the algorithm is able to diversify the set of itens fitted to the user profile and discover new relationships between the items from their attributes and their semantic relations, thereby enhancing the trade-off between accuracy and diversity of recommendation list compared with other algorithms used.