A Metric For Filter Bubble Measurement In Recommender Algorithms Considering The News Domain

Recommender systems have been constantly refined to improve the accuracy of rating prediction and ranking generation. However, when a recommender system is too accurate in predicting the users' interests, negative impacts can arise. One of the most critical is the filter bubbles creation, a situation where a user receives less content diversity. In the news domain, such effect is critical once they are ways of opinion formation. In this paper, we aim to assess the role that a specific set of recommender algorithms has in the creation of filter bubbles and if diversification approaches can decrease such effect. We also verify the effects of such an environment in the users' exposition and interaction to fake news in the Brazilian presidential election of 2018. To perform such a study, we developed a prototype that recommends news stories and presents these recommendations in a feed. To measure the filter bubble, we introduce a new metric based on the homogenization of a recommended items' set. Our results show KNN item-based recommendation with the MMR diversification algorithm performs slightly better in putting the user in contact with less homogeneous content while presenting a lower index of likes in fake news. (C) 2020 Elsevier B.V. All rights reserved.