Exploiting lod-based similarity personalization strategies for recommender systems

Abstract

Linked Open Data (LOD) is a cloud of freely accessible and interconnected datasets encompass machine-readable data. These data are available under open Semantic Web standards, such as Resource Description Framework (RDF) and SPARQL Protocol and RDF Query Language (SPARQL). One notable example of a LOD set is DBpedia, a crowd-sourced community effort to extract structured information from Wikipedia and make this information openly available on the Web. The semantic content of LOD and the advanced features of SPARQL has opened unprecedented opportunities for enabling semantic-aware applications. LOD-based Recommender Systems Recommender Systems usually leverage the data available within LOD datasets such as DBpedia to recommend items such as movies, places, books, and music to end-users. These systems use a semantic similarity algorithm that calculates the degree of matching between pairs of resources in the RDF graph, by counting the number of direct and indirect links between them, the length of the path between them, or the hierarchy of classes. Conversely, calculating similarity in RDF graphs could be difficult because each resource can have hundreds of links to other nodes. Not all of them are semantically relevant or can be applied to all resources in the graph. This can lead to the well-known matrix sparsity problem. Nevertheless, some effort has been made to select subsets of features, i.e., links, which are more helpful to computing similarity between items of a graph dataset, reducing the matrix dimension. Despite several studies in this field, there is still a lack of solutions applied to the personalization of feature selection tasks. In this context, we propose personalized strategies to improve semantic similarity precision in LOD-based Recommender Systems, including i) applying a feature selection approach to filter the best features for a particular user; ii) personalizing the RDF graph by adding weights to the edges, according to the user’s previous preferences; and iii) exploiting the similarity of literal properties as well as the links from the user model. The evaluation experiments used combined data from DBpedia and MovieLens and DBpedia and LastFM datasets. Results indicate significant increases in top-n recommendation tasks in Precision@K (K=5, 10), Map, and NDCG over non-personalized baseline similarities methods such as Linked Data Semantic Distance (LDSD) and Resource Similarity (ReSim). The results show that the strategies proposed in this work can be effective in improving semantic recommendation systems in various knowledge domains, as the solution is scalable to any LOD-based databases.