Integrating Social Environment in Machine Learning Model for Debiased Recommendation

Social data provided by social media platforms contains rich social environment information, which has been used in many ubiquitous tasks such as disaster monitoring, epidemic tracking, or stock market movement prediction. In this paper, we show that social environment information can be used to debias recommendation system. Recommendation systems are used to extract the preference of e-commerce platform users for predicting what would they see or buy next. Most recommendation systems that rely on machine learning models are currently facing the problem of bias, which occurs when the system is isolated in a platform and trained solely on past data. While various debiasing methods have been proposed, the problem remains largely unsolved. In response, we propose a recommendation model that uses real-time social data to reduce recommendation bias while also improving accuracy. The proposed model integrates social data, represented as embeddings generated by language models, into a traditional machine learning model called single value decomposition (SVD). Empirical evaluations on two pairs of real-world e-commerce plus social data datasets show that our model is superior in both recommendation accuracy and bias reduction compared to state-of-the-art debiased recommendation methods.