Unbiased Learning to Rank from user interactions

Join Textkernel for an interesting state-of-the-art Tech Talk followed by a discussion, pizza and beers!

In this talk Harrie Oosterhuis will contrast the two main approaches to unbiased learning to rank: counterfactual learning and online learning, and discuss the most recent methods from the field.

Learning to rank provides methods for optimizing ranking systems, enabling effective search and recommendation systems. Traditionally, these methods relied on annotated datasets i.e. relevance labels for query-document pairs provided by human judges. Over the years, the limitations of such datasets have become apparent, most importantly: they are expensive to create and do not necessarily reflect user preferences. Recently attention has mostly shifted to methods that learn from user interactions, as they more closely indicate user preferences. However, user interactions contain large amounts of noise and bias, learning while naively ignoring these biases can lead to detrimental results. Consequently, the current focus is on unbiased methods that can reliably learn from user interactions.

Similar events

• Mass Communication and Digital Transformation Conf - 09/09/2025 - 00.00