Offline Recommender Learning Meets Unsupervised Domain Adaptation
Year: 2019
Type: preprint
Abstract: We study the offline recommender learning problem in the presence of selection bias in rating feedback. A current promising solution to address the bias is to use the propensity score. However, the performance of the existing propensity-based methods can significantly suffer from propensity estimation bias. To solve the problem, we formulate the recommendation with selection bias as unsupervised domain adaptation and derive a propensity-independent generalization error bound. We further propose a novel algorithm that minimizes the bound via adversarial learning. Our theory and algorithm do not depend on propensity scores, and thus can result in a well-performing rating predictor without requiring the true propensity information. Empirical evaluation demonstrates the effectiveness and real-world applicability of the proposed approach.
Source: arXiv (Cornell University)
Author: Yuta Saito
Institution: Tokyo Institute of Technology
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W3036445295