Abstract: In recent years,a large number of recommendation algorithms have emerged,most of which focus on how to construct a machine learning model to give a good fit to historical interaction data.However,historical interaction data always come from observations rather than experiments in recommendation.Various biases exist in observed data,where the popularity bias is a representative one.Most approaches to dealing with popularity bias use the strategy of removing the popularity bias.But it is actually difficult for these approaches to improve the recommendation accuracy due to bias amplification causedby recommendation algorithms.Thus,the strategy of leveraging the popularity bias bothin training and inferencestagesis more applicable.Combined with the causal graph,a double bias deconfounding and adjusting(DBDA) model is proposed to rectify bias from the perspectives of both user and item.In the training stage,the adverse effects of the popularity bias are removed,and in the inference stage,a more accurate prediction of user preferences is made with the aid of the trend of popularity.Experiments are conducted on three large-scale public datasets to verify that the proposed method produces 2.48%~19.70% higher diverse evaluation metrics than the state-of-art method.

Key words: Recommender system, Collaborative filtering, Causal inference, Back-door adjustment, Popularity bias