The Need for Affective Metadata in Content -based Recommender Systems for Images

Services that provide digital images like Flickr or Facebook are attracting more and more users that generate content. For example, Flickr provides more than 3 billion of photos while Facebook provides more than 10 billion of photos (Flickr 2008 ). End users are obviously lost in these huge image databases. In order to help end users fi nd the images that are relevant for them, a recommender system could be applied (Leavitt 2006 ). State of the art content based recommender ( CBR ) systems rely on image metadata that are available along with the images. These data are stored in item profi les. Based on the users ’ past choices, user profi les are generated. These profi les contain the users ’ preferences toward various metadata values. One user might prefer images with calming content, like mountain landscapes, while another user might prefer photos depicting people. The most used metadata attribute in CBR systems is the genre . User profi les typically contain a vector of numbers representing the degree of agreeableness of the user with different genre types (Adomavicius and Tuzhilin 2005 ; Lew et al. 2006 ; Poga č nik and Tasi č 2005 ; Poga č nik et al. 2005 ). Recommender systems that exploit the genre attribute of images have reached a certain level of performance that appears to be the upper limit according to Lew et al. (2006) . In order to improve the quality of predictions of CBR systems, a better set of image attributes should be used (Lew et al. 2006 ). There have been attempts to model the users with affective metadata (Gonz á lez et al. 2004 ). Limited research has been done on the automatic extraction of induced emotive states from multimedia content (Hanjali ć 2006 ), while there is increasing effort being put in the research of emotion detection of users from multiple modalities (Picard and Daily