Propensity-Dependent Model for Unbiased Learning-to-Rank

Abstract

Most present unbiased learning-to-rank models are based on the trust bias assumption to learn a ranking policy by Inverse Propensity Scoring (IPS). The trust bias assumption improves the unrealistic noise-free assumption in the Position-Based model, but it assumes that the propensities are independent. In this paper, we improve this assumption and consider that the propensities of different positions are relevant. In particular, we model the relationship between different propensities as a Propensity-Dependent model and use both IPS estimator and the Doubly Robust estimator to learn the optimal ranking policy. Finally, we generate the dataset in a simulated study and then evaluate the model's performance.