Research And Application Of Specific Debias Algorithm For News Recommendation

The news recommendation system has been widely applied to meet the personalized news browsing needs of users and adapt to the increasing amount of news.However,personalized news recommendation algorithms are typically based on observational user implicit feedback data rather than experimental data,leading to many biases such as user selection bias and system exposure bias.Additionally,the strong timeliness of news makes user news browsing preferences highly dynamic.To address these issues,this thesis proposes two types of news recommendation models,a causal inference-based exposure-click model and an exposure-matching model that integrates temporal information.Based on these models,a news recommendation system is designed with the feasibility of the proposed models validated in practical applications.Regarding the causal inference-based exposure-click model,this thesis uses causal modeling to construct a counterfactual structure causal graph based on the exposure-click scenario in news recommendation.This is used to distinguish the effects of news exposure tendencies and user exposure preferences on user click tendencies.An improved propensity score estimation method is used to calculate users’ click propensity scores in the exposure-click causal effect calculation,and user and news information is modeled to generate news recommendation lists.Experimental results on six publicly available datasets demonstrate that the proposed model effectively alleviates exposure bias issues in news recommendation algorithms based on implicit feedback data and improves recommendation performance.Regarding the exposure-matching model that integrates temporal information,this thesis generates news exposure representations using news headlines and user’s exposure preference representations from the exposure information of the news list clicked by the user.Furthermore,the time interval information in the user’s browsing history is taken as input for modeling user news exposure preferences to represent the dynamics of user’s click preferences.The experimental results show that this model more accurately represents user’s click preferences and improves recommendation performance.Finally,based on the proposed two types of news recommendation models,combined with practical application scenarios,a news recommendation system was designed and developed to complete functions such as online news browsing,personalized news recommendation and news publishing.The feasibility of the proposed recommendation models in practical application scenarios was verified.