Taming the dynamics of recommender systems

Recommender systems can provide content that matches the interests of their users, but too often, their recommendations are redundant or overly biased towards a few categories of content. This is known to degrade the user experience, and affect both user-side and business-side metrics in the long run, and it is suspected to have detrimental consequences on certain societal concerns such as fairness and radicalization pathways. Part of the reasons why this happens is the fact that recommender systems do not account for the effect they have on people using them.
In this thesis, I investigate recommendation models and agents that are able to acknowledge the consequences of deploying them to a live recommendation platform, and can therefore correctly optimize the target metrics in the long-run. In particular, I distinguish two types of distribution shifts, i.e., of changes in user-system interactions due to the deployment of recommender systems: (i) algorithmic filtering in the selection and presentation of content that results in interaction logs that are biased towards certain items, and (ii) the effect of recommendations on a user's behavior and preferences (e.g., boredom, new interests, biased worldviews, ...) when multiple recommendations are made sequentially. Because of these effects of recommender systems on their users, recommendation is an inherently dynamic and interactive task.