Title：Unbiased Learning to Rank from User Interactions
Location：N3-320
Time：15:00-16:00, 21st May 2019
Speaker：Prof. Dr. Maarten de Rijke, University of Amsterdam
Abstract:
Learning to rank is a core ingredient of search engines, recommender systems, and digital assistants. Supervised approaches to learning to rank are limited in that Annotations often disagree with user preferences. User interactions solve this problem but bring noise and biases. Counter-factual approaches allow for unbiased learning to rank: if an accurate user model can be learned, we can adjust for biases; we only require randomization to infer a user model. In contrast, online approaches allow for unbiased and responsive learning to rank, as they immediately adapt to user behavior and perform randomization at each step, though limited. Clearly, different situations suit different approaches. We report on recent advances and on recent findings on comparing the two families of approaches.
