Machine Learning Seminar: Unbiased Learning-to-Rank with Biased Data
Logged user interactions are one of the most ubiquitous forms of data available, as they can be recorded from a variety of systems (e.g., search engines, recommender systems, ad placement) at little cost. While it is clear that interactions (e.g. clicks) reveal some information about the user's preferences, this observable feedback is biased by the actions of the system (e.g. presented search ranking). This creates challenges when using log data for learning.
To overcome these challenges, this talk explores how to use techniques from causal inference and missing-data analysis to account for selection biases in Learning-to-Rank (LTR). First, I will derive debiased training objectives based on inverse-propensity-score (IPS) weighting estimators, as well as their implementation in practical methods for LTR. Second, I will propose a new class of contextual propensity models for rankings and an associated estimator that does not require disruptive randomized interventions. Beyond their theoretical justification, both the LTR methods and the propensity estimators are evaluated in simulation studies and in operational systems.