In the era of information overload, social recommender systems, which leverage the emergence of online social networks to perform personalized content filtering based on user preferences, have proven successful. However, social recommender models not only exhibit a well-known bias toward popular items but also have a social popularity bias that is often overlooked in existing research. Both biases can lead the model to learn inaccurate user representations, ultimately compromising the diversity and accuracy of recommendations. This paper focuses on integrating social networks into recommendations in an unbiased way. First, a new causal graph is proposed to understand how item and social popularity affect user representation and how user consistency preferences affect ranking scores. Next, to eliminate the adverse effects of popularity bias, we explore how to leverage backdoor adjustments to learn unbiased user representations and obtain accurate ranking scores through a counterfactual reasoning strategy. Finally, using the backdoor adjustment operator and the counterfactual reasoning strategy as key components, a causal inference-based social graph neural network is proposed. Evaluation results on four real-world data sets show that our proposed model surpasses state-of-the-art methods in recommendation accuracy and diversity. History: Accepted by Ram Ramesh, Area Editor for Data Science & Machine Learning. Funding: This work was supported by the National Natural Science Foundation of China Grants 72271195 and 72472128. Supplemental Material: The software that supports the findings of this study is available within the paper and its Supplemental Information ( https://pubsonline.informs.org/doi/suppl/10.1287/ijoc.2024.0682 ) as well as from the IJOC GitHub software repository ( https://github.com/INFORMSJoC/2024.0682 ). The complete IJOC Software and Data Repository is available at https://informsjoc.github.io/ .
Xu et al. (Fri,) studied this question.