Philipp Hager

In this project, we explore deep probabilistic models for learning ranking systems with user feedback (as is common in web search or movie/song/hotel recommendation). Feedback like clicks, purchases, or views is commonly used to optimize machine learning algorithms. Yet, they are often a (statistically) biased signal of user preference since users can only interact with items previously shown to them. The field of unbiased learning-to-rank (ULTR) deals with these biases in the context of optimizing machine learning models for ranking.

Our first work compares bias mitigation using traditional probabilistic models with more recent methods from causal inference for ULTR (ECIR 2023). Our second work tackles challenges in the offline evaluation of probabilistic models before deploying them (SIGIR 2023). Recently, we investigated the real-world effectiveness of ULTR methods, as the field previously primarily evaluated methods through simulation experiments. We found large discrepancies between simulation and real-world data and formulated multiple interesting research avenues (SIGIR 2024). In upcoming work, we explore the theoretical conditions for training certain classes of deep probabilistic models on click data.