Towards more impactful recommender systems research

A number of ways exist in which a recommender system can have an impact on users and business. However, only a small number of them can be reasonably addressed with today’s predominant and narrow research approach based on offline experimentation and accuracy measures. It sometimes even stands to question if small increases in prediction accuracy will actually lead to a better system in any of the ways in which a recommender system can impact users and create value. We therefore argue for a more impact-oriented approach to research in the field of recommender systems. With such a refocused lens, we hope that the corresponding research results are also more impactful and relevant in reality. To foster such research, we present in this work a first taxonomy describing the various facets to consider when developing impact-oriented research, ranging from the expected value of a recommender for different stakeholders to the potential risks that come with such applications.