Multiple Robust Learning for Recommendation

In recommender systems, a common problem is the presence of various biases in collected data, which deteriorates generalization ability recommendation models and leads to inaccurate predictions. Doubly robust (DR) learning has been studied many tasks RS, with advantage that unbiased can be achieved when either single imputation or propensity model accurate. this paper, we propose multiple (MR) estimator take candidate achieve unbiasedness. Specifically, MR any models, linear combination these Theoretical analysis shows proposed an enhanced version DR only having model, smaller bias. Inspired by error bound MR, further novel approach stabilization. We conduct extensive experiments on real-world semi-synthetic datasets, demonstrates superiority over state-of-the-art methods.