Causal Disentangled Sentiment Debiasing for Recommendation

Recommender system usually suffers from bias problem. To deal with this problem, recommendation debiasing has received much attention recently. In line with this trend, our work aims to shed light on a newly identified bias known as sentiment bias - the divergence in recommendation performance between positive users/items and negative users/items. By investigating this aspect, we hope to contribute to understanding and mitigating the sentiment bias. Existing efforts usually focus on utilizing regularization strategies or causal interventions for eliminating the bias. However, these methods fall short of decomposing the underlying user preference or sentiment bias, which limits their ability to significantly enhance recommendation performance. In this study, we address the issue of sentiment bias by incorporating disentanglement and attention networks from a causal perspective. We construct a causal graph to capture the cause-effect relationships in the recommender system, specifically considering the role of sentiment polarity expressed in review text as a confounding factor between user/item representations and observed ratings. To mitigate the negative impact of sentiment bias, we propose a disentangled framework that learns representations where user preferences and sentiment bias are structurally disentangled. Furthermore, we use an attention network to extract the expected sentiment of the candidate items from the users historical behaviors. Extensive experiments conducted on five benchmark datasets validate the effectiveness of disentanglement in removing sentiment bias.