A hybrid approach for artwork recommendation

Museums usually exhibit thousands of artworks, and nowadays, they often have their collections online for visitors. In these collections, the curators are responsible for organizing the artworks seeking a delicate balance between emotion and reason. Given an initial artwork, however, a visitor is likely to select and admire a set of related artworks that match her interests. This setting can be seen as a recommendation problem in the art domain. Although image recommendation systems have been previously developed, considering the artwork nature is a fundamental aspect when designing a recommender system in this domain. Thus, we propose a hybrid recommendation approach that combines deep autoencoders with a social influence graph in order to capture the visual aspects and context of artworks (represented by images). These mechanisms inform the generation of rankings of related artworks. In this context, we report on a case-study with a group of art experts who assessed the rankings of artworks recommended by our approach. Although preliminary, the results showed a better precision than traditional strategies based solely on image features or metadata. Furthermore, the recommendations exhibited diversity properties, avoiding typical over-specialization problems of content-based techniques. •We propose an artwork recommendation approach that combines deep autoencoders with a social influence graph.•Visual-content features and context features are jointly used to improve the recommendation.•The performance of the proposed approach is evaluated with a group of art experts and compared against a content-based approach.