Research On Personalized News Recommendation Based On Knowledge Enhancement And Preference Fusion

The update speed of news data in the Internet is very fast.In the face of the huge amount of news in the Internet,how to quickly and accurately push the news they are interested in to users has become an urgent problem for news recommendation.When browsing news,the news titles tend to be more attractive to users,which indicates that users are more sensitive to titles.Similarly,keywords in news titles often reflect the subject of the news.So in personalized news recommendation,titles have important research significance.Unlike the general commodity attributes,news contains rich text and event information,which can supplement the content of news titles to a certain extent.Therefore,how to accurately recommend users based on these known information has become the main direction of news recommendation research.Furthermore,the diversity of user interests in different news categories may lead to changes in interests,which impact news recommendation.Currently,most news recommendation models directly consider users’ browsing history as their interests,but latent interests are difficult to reflect through explicit behaviors.Moreover,user interests are dynamic and capturing the dynamics of user interests is crucial for feature representation.Although interests influence each other,each interest has its own evolutionary process.Neglecting the correlation between interests can affect the effective representation of user interests,thereby impacting the accuracy of news recommendation.To solve the above problems,this thesis mainly completes the following work:(1)Research on a news recommendation model based on knowledge enhancement and attention mechanismThis thesis has improved a new model-news recommendation model based on knowledge augmentation and attention mechanism(KEAN).The KEAN leverages knowledge graph and graph attention networks to capture entities and their contextual information in news titles for feature representation.It also introduces multi-layer attention networks to model user’s current interests and uses attention mechanism to calculate the impact of user’s click behavior on candidate news,thus modeling user’s current interests.Experimental results have shown that the proposed KEAN model shows good performance on Adressa-1week and Adressa-10 weeks datasets.Compared with the best performing baseline model(CVAR),the KEAN model improves F1 by 1.2%and AUC by 1.6% on the Adressa-1 week dataset;On the Adressa-10 weeks dataset,there was a 1.4% increase in F1 metrics and a 1.1% increase in AUC metrics.Compared with the news recommendation model in the past three years,the F1 index has increased by 1.1% to 2.5%,and the AUC index has increased by 1.2% to 2.7%.This also indicates that the richness of news entities and the mining of user interests in the model can fully improve the accuracy of recommendations.(2)Research on graph neural news recommendation model based on deep interest networksOn the basis of the existing KEAN model,in order to solve the problems of user interest transformation and high-order representation of candidate news,this thesis improves a new news recommendation model,namely the News Recommendation Model Based on Enhanced User Interest(News EUI).To fully utilize the interaction information between users and news,and obtain comprehensive high-order encoding of user browsing information and candidate news,this thesis first constructs a user-news interaction graph by explicitly representing the connectivity between news and users.Then,multi-layer graph convolutional networks(GCN)are employed to learn on the graph structure,resulting in stable and long-term representations of user interests and high-order representations of candidate news.Furthermore,considering the impact of changes in user interests on recommendation results,this thesis introduces gated recurrent units(GRU)and self-attention updated gates in GRU(SAUGRU)to extract short-term user interests,enriching the model’s representation of user preferences in data-sparse scenarios and improving the interpretability of recommendation results.Experimental results on Adressa-1week and Adressa-10 weeks datasets demonstrate that the News EUI model outperforms the KEAN model in terms of performance.Particularly,the performance improvement is more significant in scenarios with limited data,validating the effectiveness of using interaction graphs to address data sparsity.Compared with other advanced news recommendation models in the past three years,on the Adressa-1 week dataset,the improvement in F1 indicators is 1.07% to 3.79%,and the improvement in AUC indicators is 1.37% to 3.72%;On the Adressa-10 weeks dataset,the improvement range on the F1 indicator is 1.33%～4.92%,and the improvement range on the AUC indicator is 1.31%～2.74%.