Exploiting Linked Open Data in Cold-start Recommendations with Positive-only Feedback

In recommender systems, user preferences can be acquired either explicitly by means of ratings, or implicitly --e.g., by processing text reviews, and by mining item browsing and purchasing records. Most existing collaborative filtering approaches have been designed to deal with numerical ratings, such as the 5-star ratings in Amazon and Netflix, for both rating prediction and item ranking (a.k.a. top-N recommendation) tasks. In many e-commerce and social network sites, however, user preferences are usually expressed in the form of binary and unary (positive-only) ratings, such as the thumbs up/down in YouTube and the likes in Facebook, respectively. Moreover, in these cases, the well-known problem of cold-start --i.e., the scarcity of user preferences-- is highly remarkable. To address this situation, we explore a number of graph-based and matrix factorization recommendation models that jointly exploit user ratings and item metadata. In this work, such metadata are automatically obtained from DBpedia --the queriable and structured version of Wikipedia which is considered as the core knowledge repository of the Linked Open Data initiative--, and the models are evaluated with a Facebook dataset covering three distinct domains, namely books, movies and music. The results achieved in our experiments show that the proposed hybrid recommendation models, which exploit rating and semantic data, outperform content-based and collaborative filtering baselines.