Doubly Robust Collaborative Targeted Learning for Debiased Recommendations

Abstract

In recommender systems, the feedback data received is always missing not at random (MNAR), which poses challenges for accurate rating prediction. To address this issue, many recent studies have been conducted on the doubly robust (DR) method and its variants to reduce bias. However, theoretical analysis shows that the DR method has a relatively large variance, while that of the error imputation-based (EIB) method is smaller. In this paper, we propose DR-TMLE that effectively captures the merits of both EIB and DR, by leveraging the targeted maximum likelihood estimation (TMLE) technique. DR-TMLE first obtains an initial EIB estimator and then updates the error imputation model along with the bias-reduced direction. Furthermore, we propose a novel RCT-free collaborative targeted learning algorithm for DR-TMLE, called DR-TMLE-TL, which updates the propensity model adaptively to reduce the bias of imputed errors. Both theoretical analysis and experiments demonstrate the advantages of the proposed methods compared with existing debiasing methods. * Contributed equally † Contact Author Preprint. Under review.

Stabilized Doubly Robust Learning for Recommendation on Data Missing Not at Random

“…L ideal (φ) can be regarded as a benchmark of unbiased loss function, even though it is infeasible due to the missingness of {r u,i (1) : o u,i = 0}. As such, a variety of methods are developed through approximating L ideal (φ) to address the selection bias, in which the propensity-based estimators show the relatively superior performance [25,31,6,5,37], and the IPS and DR estimator is given as…”

“…A doubly robust joint learning (DR-JL) method [31] was proposed by combining the IPS and EIB approaches. Subsequently, strands of enhanced joint learning methods were developed, including more robust doubly robust (MRDR) method [6], multi-task learning [40], collaborative targeted learning [37], and uniform data-aware methods [3,12,5,32] that aimed to seek better recommendation strategies by leveraging a small uniform dataset. [4] reviewed various biases in RS and discussed the recent progress on debiasing tasks.…”

“…The proposed method builds on the stabilized average caual effect estimation approaches in causal inference. [16,37] summarized the limitations of doubly robust methods, including unstable to small propensities [36], unboundedness [30], and large variance [29]. These issues inspired a series of stabilized causal effect estimation methods in statistics [10,2,30,16].…”