Learning-to-rank systems often utilize user-item interaction data (e.g., clicks) to provide users with high-quality rankings. However, this data suffers from several biases, and if naively used as training data, it can lead to suboptimal ranking algorithms. Most existing bias-correcting methods focus on position bias, the fact that higher-ranked results are more likely to receive interaction, and address this bias by leveraging inverse propensity weighting. However, it is not always possible to accurately estimate propensity scores, and in addition to position bias, selection bias is often encountered in real-world recommender systems. Selection bias occurs because users are exposed to a truncated list of results, which gives a zero chance for some items to be observed and, therefore, interacted with, even if they are relevant. Here, we propose a new counterfactual method that uses a two-stage correction approach and jointly addresses selection and position bias in learning-to-rank systems without relying on propensity scores. Our experimental results show that our method is better than state-of-the-art propensity-independent methods and either better than or comparable to methods that make the strong assumption for which the propensity model is known.

Click data collected by modern recommendation systems are an important source of observational data that can be utilized to train learning-to-rank (LTR) systems. However, these data suffer from a number of biases that can result in poor performance for LTR systems. Recent methods for bias correction in such systems mostly focus on position bias, the fact that higher ranked results (e.g., top search engine results) are more likely to be clicked even if they are not the most relevant results given a user’s query. Less attention has been paid to correcting for selection bias, which occurs because clicked documents are reflective of what documents have been shown to the user in the first place. Here, we propose new counterfactual approaches which adapt Heckman’s two-stage method and accounts for selection and position bias in LTR systems. Our empirical evaluation shows that our proposed methods are much more robust to noise and have better accuracy compared to existing unbiased LTR algorithms, especially when there is moderate to no position bias.