OLIVEIRA, R. S.; http://lattes.cnpq.br/6500030283495313; OLIVEIRA, Ricardo Santos de.
Resumen:
Many success ful recommendation approache sare based on the optimization of some explicit utility function defined in term sof them is fit between the predicted and the actual item sof the user. Al though effective, this approach may lead to recommendations that are relevant But obvious and uninteresting. Many approaches investigate this problem by trying to avoid recommendation lists in which item sarevery similar toeachother (akadiversification) with respect to some aspect of the item. However, users may have very diferente preferences concerning what aspects should be diversified and whats hould match their past/current preferences. In this work, we take this in to consideration by proposing solutions base don multi objective optimization for generating recommendation lists featuring the optimal balance between the aspects thats hould be held fixed (maximize similarity with users actual items) and the one sthat should be diversified (minimize similarity with other item sin the recommendation list). In order to evaluate the proposed models, simulations were carried out using real Last.fm Data sets, together with metadata about there commended items. Evaluations were carried out both offline, that is, using the entire user history to produce there commendation, and online, where recommendations are generated along the timeline, takingin to account changes in the user’s interests that should be reflected in the recommendation, which resembles the actual application of are commendation system. The results obtained in the experiments
demonstrate the effective ness of the models proposed both offline and online, compared to state-of-the-art approaches in the literature with similar proposals.