Recent advances in learning ranking models from click data

What is it about?

Algorithms that help us search the web or recommend content such as movies, songs, or hotels are prevalent in our everyday life. A crucial part of these systems are machine learning models that learn how to rank items in order of relevance to a user, often from implicit user feedback (such as clicks, views, watch time, etc.). While often readily available, implicit feedback is typically a noisy and biased signal of user preference. Factors such as an item's position on the screen, an item's visual presentation, or the user's trust in a search engine to deliver relevant results might introduce bias into implicit user feedback, biases that the field of unbiased learning to rank seeks to mitigate. In this tutorial at the 2023 ACM SIGIR conference, we introduce the core concepts of unbiased learning to rank, an overview of recent advances in its foundation, and connect to the related field of fairness in ranking.

Featured Image

Photo by Jonas Leupe on Unsplash

Photo by Jonas Leupe on Unsplash

Why is it important?

Since the last tutorial, the field of unbiased learning to rank matured significantly. Our tutorial covers new estimation techniques, bias mitigation beyond position bias, and explores the connection to fairness in ranking applications. This overview will benefit both academic researchers and industry practitioners interested in developing new ULTR solutions or utilizing them in real-world applications.