Pull–push: a measure of over- or underpersonalization in recommendation

Abstract A recommender system imposes differences between users, by presenting to them different recommendation lists, which they respond to, resulting in “reaction” lists. Comparison of the and reaction lists can indicate user states. Users approve imposed difference, end up narrowing difference (pulling each other closer) consuming more items common or enlarge (pushing further apart) not common. When users do differences, are either a push state (implicitly disapproving under-personalization) pull over-personalization). We offer pull–push metric quantify magnitude push—measures disapproval of, respectively, over-personalization under-personalization. Application on simulated datasets shows that away disjoint sets closer identical sets. On real-world datasets, we find particular was under-personalizing its recommendations. show how be merged with another personalization come measure potential for improvement discuss relationship popularity bias.