Counterfactual Inference for Eliminating Sentiment Bias in Recommender Systems

TL;DR

This work tackles sentiment bias in review-based recommender systems by leveraging counterfactual inference. It introduces a causal graph with sentiment as a mediator and uses natural direct effect to separate direct user/item influences from sentiment-driven indirect effects, applying debiased inference at prediction time. The model integrates sentiment-aware branches and auxiliary losses within a neural collaborative filtering backbone, achieving strong MSE performance while substantially reducing user- and item-level sentiment bias across multiple datasets. The findings demonstrate the practicality and effectiveness of counterfactual approaches for fairer, more robust recommendations in the presence of emotionally charged reviews.

Abstract

Recommender Systems (RSs) aim to provide personalized recommendations for users. A newly discovered bias, known as sentiment bias, uncovers a common phenomenon within Review-based RSs (RRSs): the recommendation accuracy of users or items with negative reviews deteriorates compared with users or items with positive reviews. Critical users and niche items are disadvantaged by such unfair recommendations. We study this problem from the perspective of counterfactual inference with two stages. At the model training stage, we build a causal graph and model how sentiment influences the final rating score. During the inference stage, we decouple the direct and indirect effects to mitigate the impact of sentiment bias and remove the indirect effect using counterfactual inference. We have conducted extensive experiments, and the results validate that our model can achieve comparable performance on rating prediction for better recommendations and effective mitigation of sentiment bias. To the best of our knowledge, this is the first work to employ counterfactual inference on sentiment bias mitigation in RSs.

Figures (6)

• Figure 1: An illustration of sentiment bias.
• Figure 2: Causal Graph for (a) Traditional user-item matching paradigm. Node $U$ represents the user variable, which refers to the user profile, including review records and interaction history. Node $I$ is the item variable that contains item data, review records and the interaction history. Node $Y$ is the rating variable, which is the output of RRSs. Edge $U \rightarrow Y$ represents the direct effect from user representation to rating. Edge $I \rightarrow Y$ represents the direct effect from item representation to rating. (b) Incorporating sentiment bias; Node $S$ is the sentiment variable that represents the sentiment in the reviews. Edge $U \rightarrow S \rightarrow Y$ and $I \rightarrow S \rightarrow Y$ represent the indirect effect on rating originating from user and item, respectively, with $S$ as the mediator variable. Edge $S \rightarrow Y$ represents the sentiment bias recently proposed by lin2021mitigating, which reveals the divergence of recommendation performance between positive user/item and negative user/item. (c) Counterfactual inference. Grey nodes are in the reference state pearl2009causality, for example, $u^*$ means $U=u^*$.
• Figure 3: Our proposed RRSs pipeline is based on counterfactual inference. This pipeline consists of (a) embedding computation, (b) capturing sentiment, (c) modelling the user-item interaction, and (d) rating prediction.
• Figure 4: Boxplots on mean MSE comparison before and after debias(RQ3).
• Figure 5: Difference between before and after debias in rating distribution shift(RQ3).