Diversity of What? Different Conceptualizations of Diversity in Recommender Systems

Research Outline

The Diversity in Recommender Systems Project centres around the multitude of definitions that apply to the concept of diversity. Sanne Vrijenhoek explores how to navigate this discrepancy by conducting an interview series with industry practitioners. Additionally, she challenges the preconceptions about recommender systems in the NORMalize workshop series.

Diversity is a heavily contested and ambiguous concept. In popular discourse, diversity usually refers to the variation of human characteristics, often in the context of identity. In biology, diversity is a qualifier for a healthy ecosystem, while in media studies it might indicate a plurality of opinions or sources. Although diversity seems to intuitively mean similar things, the operationalization of the term is unique to each field.

In recommender systems, diversity is understood as a countermeasure to monotony. In news recommendation for example, it can work against the creation of filter bubbles and expose users to unexpected things, while still accommodating their information needs. Furthermore, for machine learning, diversity is important in a social context, as algorithms are often trained on biased datasets and are not representative of all groups in society. Here, not accounting for diversity may have harmful consequences.

These many different interpretations of diversity are a fundamental challenge to the practical development of recommender systems. One could argue that diversity is, in fact, an essentially contested concept meaning that it is open for discussion and debate and that we are unlikely to reach consensus on its meaning. Striving for agreement or a clear definition may lead to a standstill and hinder progress, as it is unclear what a good operationalization or implementation would look like.

Instead, we may need a more flexible operationalization of diversity that is capable of reflecting the nuances and requirements of the domain it is deployed in.

Beyond Academia

It is of critical importance that these discussions do not remain limited to academic discourse. A workshop series is currently being prepared with the AI, Media and Democracy Lab. It aims to help media organizations conceptualize and formalize the type of diversity relevant to their system. The organizations can go through all the different parts of the workshop series in a self-guided manner: from defining the goal of the recommender system, to what type of diversity corresponds to that goal, to deciding what can and should be measured, to expressing and communicating the results of the metric in a meaningful way.