What if? Interaction with Recommendations

Showing users recommended content has become a prominent way of integrating algorithmic decision-making in everyday intelligent applications (e.g. recommendations of films, music, news, routes). In this context, the research community has identified What if? questions as an approach for users to investigate and question such recommendations – yet many current applications seem limited in practically supporting this. We present a set of example GUIs and interaction techniques currently used in everyday recommendation systems in practice (e.g. Grammarly, Apple Music, Google Maps). Based on these example cases, we discuss possible UI extensions to explicitly supportWhat if? interactions. From our analysis and reflection emerges the general approach of treating decision variables as a “first-class citizen” in UIs: We propose to 1) represent a recommended item’s decision variables in the user interface (and not just the item itself), and 2) to enable direct manipulation of these decision variables forWhat if? explorations.