Ratings in, rankings out. Keep it simple, they said. But we need more than that

Among the many viable research questions in the field of recom-mender systems, a frequently addressed problem is to accurately predict the relevance of individual items to users, with the goal of presenting the assumedly most relevant ones as recommendations. Typically, we have users' (explicit or implicit) ratings as input and rankings of items as output. Complex enough, yet too simplistic to reflect reality and indeed meet the various demands in practice. We have learned that "context matters". But what does it mean? What is the context that matters? And how do we get the relevant signals? It is more than what we currently ascribe to and reflect in what we call "context-aware recommender systems". Let's have a view to related fields that deal with context as deeply complex input. And on the output side, we have individual items and also item bundles, complementaries, sequences, repeated recommendations, etc. What do we actually want to present? And how? For who? And why? A ranked list as output may seem like an appropriate one-size-fits-all solution, does it? In this talk, I will reflect on the complexity of our research field, reach out to related fields such as context-aware computing and pervasive advertising for inspiration, and I will raise a lot of questions that have yet to be answered.