A new approach to calibrating recommendations to user interests
Users’ interests are multi-faceted and representing different aspects of users’ interest in their recommendations is an important factor for recommender systems to help users navigate more quickly to items or content they may be interested in. This property is often referred to as the calibration problem and has achieved significant attention recently. Calibration is particularly important given that a sole optimization towards accuracy can often lead to the user’s minority interests being dominated by their main interests, or by a few overall popular items, in the recommendations they receive. In this work, we propose a novel approach based on a minimum-cost flow through a graph for generating accurate and calibrated recommendations.
Calibration in Recommender Systems
Recommender systems often optimize for the most relevant items to the user. Suppose a user has listened to a lot of pop music, some jazz music and also some podcasts. The recommender algorithm may return a list of recommendations that are all pop music (ignoring other types of music and also ignoring podcasts) or a list containing all music (ignoring podcasts) as those might be the most relevant to the user according to the objective function. Would such a recommendation list be useful to the user? Calibration in recommendation refers to the fact that a good recommendation list should reflect various aspects of the user’s interest ideally in the right proportion.
To continue reading this article, click here.