Filter bubbles created by collaborative filtering algorithms themselves, fact or fiction? An experimental comparison

Abstract

By recommending only, or mostly, items close to our supposedly known tastes, a recommender system could lead to the narrowing of the diversity/novelty of the recommended list, leading to the creation of so-called filter bubbles. Authors however disagree on the real existence of filter bubbles and, if real, on the responsibility of algorithms in the process of their creation. This paper aims to experimentally analyze the potential creation of such filter bubbles by testing various collaborative filtering algorithms on their intrinsic propensity of changing the diversity/novelty of recommended items, in configurations where users behave, with various intensities, in a way that could influence the evolution in diversity/novelty. Results show that both human and algorithmic components could have an impact on diversity/novelty, depending on the type of algorithm.