# Balancing Unobserved Confounding with a Few Unbiased Ratings in Debiased Recommendations

**Authors:** Haoxuan Li, Yanghao Xiao, Chunyuan Zheng, Peng Wu  
**arXiv:** [2304.09085](https://arxiv.org/abs/2304.09085)  
**Date:** 2023-04-17  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.09085v1/

## Key points

- Recommender systems are important to address information overload
- Biases in large-scale observational data can result in sub-optimal prediction performance
- Unbiased ratings obtained from randomized controlled trials or A/B tests can correct biases, but existing methods have limitations
- Trovato et al. propose a theoretically guaranteed model-agnostic balancing approach that can be applied to any existing debiasing method
- The proposed approach makes full use of unbiased data by correcting model parameters learned with biased data and adaptively learning balance coefficients of biased samples for further debiasing
- Various approaches developed for achieving unbiased learning such as inverse propensity score (IPS) strategy, doubly robust (DR) loss, multiple robust learning method and adversarial learning method are discussed
- Challenges associated with addressing unmeasured confounding in causal inference are highlighted and sensitivity analysis and auxiliary information-based methods are discussed for obtaining unbiased causal effect estimators.
- The proposed balancing optimization training objective is demonstrated through extensive real-world experiments on two datasets to combat the effects of unobserved confounding and model misspecification.
- Leveraging a small uniform dataset can achieve exact unbiasedness in debiased recommendations.
- Future work includes deriving theoretical generalization error bounds for balancing approaches and exploring more effective ways to enhance debiasing performance using unbiased ratings.

## Summary

Recommender systems have become an essential tool to address information overload. However, the presence of various biases in large-scale observational data often results in sub-optimal prediction performance. To combat this issue, recent works proposed using unbiased ratings obtained from randomized controlled trials or A/B tests to correct the parameters of propensity or imputation models trained on biased datasets. However, existing methods fail to obtain accurate predictions in the presence of unobserved confounding or model misspecification. To address these limitations, Trovato et al. propose a theoretically guaranteed model-agnostic balancing approach that can be applied to any existing debiasing method. The proposed approach makes full use of unbiased data by alternatively correcting model parameters learned with biased data and adaptively learning balance coefficients of biased samples for further debiasing. The paper also discusses various approaches developed for achieving unbiased learning such as inverse propensity score (IPS) strategy, doubly robust (DR) loss, multiple robust learning method and adversarial learning method. Additionally, the authors highlight the challenges associated with addressing unmeasured confounding in causal inference and discuss sensitivity analysis and auxiliary information-based methods for obtaining unbiased causal effect estimators. The proposed balancing optimization training objective is demonstrated through extensive real-world experiments on two datasets to combat the effects of unobserved confounding and model misspecification. The study shows that leveraging a small uniform dataset can achieve exact unbiasedness in debiased recommendations. Overall, this paper presents a novel approach to combating biases in recommender systems by fully exploiting the unbiasedness property of a small uniform dataset while addressing unobserved confounding and model misspecification. Future work includes deriving theoretical generalization error bounds for balancing approaches and exploring more effective ways to enhance debiasing performance using unbiased ratings.