Improving accuracy by reducing the importance of hubs in nearest-neighbor recommendations

Van Parijs, Clémentine [UCL] Fouss, François [UCL]

A traditional approach for recommending items to persons consists of including a step of forming neighborhoods of users/items. This work focuses on such nearest-neighbor approaches and, more specically, on a particular type of neighbors, the ones frequently appearing in the neighborhoods of users/items (i.e., very similar to many other users/items in the data set), referred to as hubs in the literature. The aim of this paper is to explore through experiments how the presence of hubs aects the accuracy of nearest-neighbor recommendations.