Uncertainty Estimation Based Doubly Robust Learning for Debiasing Recommendation

In recommender systems, the interactions between users and items are sparse, which means most of the interactions between users and items are missing. More importantly, the loss of these interactions is non-random, which leads to selection bias and makes it difficult to predict the ratings accurately. Most recent methods combine the error-imputation-based (EIB) estimator and the inverse-propensity-score (IPS) estimator to overcome this problem. However, these methods may result in even higher variance because of the inaccurate error imputation, so we propose to make an uncertain estimate of error imputation and eliminate imputed error with excessive variance. We developed a novel framework for recommender system debiasing by estimating the inaccuracy of imputed errors. Specifically, we use dropout technology to calculate the variance of imputed error. Then, we filter imputed errors based on variance. Imputed errors with higher variance are less likely to be reserved which improves the credibility and quality of imputed errors. In this way, the performance of the estimator can be further improved. We conduct extensive experiments on two real-world datasets which demonstrate that our proposed framework has strong stability and outperforms the state-of-the-art.