Exploiting item co-utility to improve recommendations

Abstract

In this work we show that the feedback items receive in a recommender system can be exploited in order to improve recommendations. If two items consistently receive positive feedback together, the selection of one of them should influence the other's selection. We then present different ways of quantifying such influence, all of them computationally cheap and independent of content information about either items or users. Afterwards, we plug this information in a version of the Facility Dispersion Problem, to which our problem reduces, and present different algorithms for solving it. We compare our approach with competitive baselines and obtain expressive gains in the utility of recommendations. We also show that our approach is scalable and does not affect recommendations' diversity.