Causal Debias Method for Multiple Confounding Factors in Recommendation
How to eliminate the bias problem existing in recommendation is a major challenge for recommender systems to improve their effectiveness.The existing recommendation models only fit the data,and eliminating bias based on correlation may be affected by spurious correlations.Eliminating bias based on causal relationships is difficult to extract more comprehensive causal relationships due to the complex scene.Therefore,it is necessary to consider the causal relationships between various factors as much as possible and eliminating bias issues.We study the causal relationship of the two confounding factors in the recommendation process,namely,the dis-tribution imbalance of user behavior in item classification and item category popularity.A de-confound method is proposed that considers multiple factors to effectively eliminate the bias problem in recommendation.Firstly,the causal relationship between variables in the rec-ommendation is analyzed and a causal graph is constructed.Secondly,the false correlation as well as the bias caused by the two confounding factors are eliminated using front-door adjustment and back-door adjustment.Finally,the proposed method is applied to neural factorization machines,and experimentally validated on two publicly available datasets,the results show that the proposed method has varying degrees of improvement compared to the current optimal method.
国家科技期刊平台
NSTL
万方数据
