Author Description

"This thesis describes a machine learning-based personalization flow for streaming platforms: we match users and content like video or music, and monitor the results. We find that there are still many open questions in personalization and especially in recommendation. When recommending an item to a user, how do we use unobservable data, e.g., intent, user and content metadata as input? Can we optimize directly for non-differentiable metrics? What about diversity in recommendations? To answer these questions, this thesis proposes data, experimental design, loss functions, and metrics. In the future, we hope these concepts are brought closer together via end-to-end solutions, where personalization models are directly optimized for the desired metric."--