Joint Training of Propensity Model and Prediction Model via Targeted Learning for Recommendation on Data Missing Not at Random

Hao Wang

Keywords: Missing Not At Random, Target Learning

Abstract: Recommender systems (RS) help to capture users' personalized interests and are increasingly important across social media, e-commerce, and various online applications. Since users are free to choose items to rate, the collected ratings are a missing-not-at-random subset of all user-item pairs, and there is a systematic distributional shift between observed and unseen ratings. The bias caused by the distributional shift is called the selection bias. There have been emerging quantities of methods to address the selection bias. The error-imputation-based (EIB), inverse propensity score (IPS), and doubly robustness (DR) try to improve the prediction accuracy by introducing the imputation model and propensity model. On top of these fundamental methods, some enhanced methods are proposed to achieve smaller bias and variance. However, most of these methods cannot achieve the nonparametric efficiency for estimating the prediction model or lack theoretical guarantees for the robustness and efficiency. To bridge this gap, this paper uses a neural-network-based architecture to model the propensity and prediction model and jointly train the two models with a target learning approach. Specifically, we add a targeted regularization that guides the optimization in the most efficient direction. Experiments on three widely used real-world datasets show the effectiveness of our method.

Submission Number: 29