| In the era of digitalization and informatization,though abundant news and Information have met people’s growing cultural needs,they have burdened people with excessive filtering.News recommendation systems have emerged as an important information filtering tool,which not only helps users quickly and accurately find the most interesting news content,improve user experience and satisfaction,but also helps news media increase subscription,user activity,and advertising revenue through personalized recommendation mechanisms,enhancing the commercial value and market competitiveness of the media.Thereby,research on news recommendation systems can also promote information exchange and dissemination,and promote social progress and development.News recommendation researches mainly involves two key tasks: News modeling,which aims to convert the raw news information into vector representations with semantic meaning;User modeling,which aims to understand user behaviors and needs in order to achieve personalized recommendation.Currently,both two tasks suffer from insufficient representation capabilities,which make it difficult for existing methods to fully utilize news semantics and effectively capture users’ dynamic preferences.To address these issues,this thesis aims to further improve the quality of news recommendation by conducting research from two levels:(1)improving the network’s ability to extract news semantics,and(2)improving the network’s ability to mine user interests.(1)A personalized news recommendation method based on text heterogeneous modeling and historical local activation.The method’s core includes the Text Heterogeneous Modeling(THM)module and the Historical Local Activation(HLA)module.In order to address the issue of insufficient semantic understanding caused by using the same network structure to encode title and content texts,the network structure design of the THM news encoding module fully considers the differences between the title and content text,and uses different deep network structures to learn their representation vectors according to semantic characteristics.This “made-to-measure”approach improves the representation effect of news.In order to address the issue of capturing fine-grained interests in the user’s history,the HLA user encoding module divides the user’s history into slices,learning the user’s preference representation vectors in different time periods to more accurately describe the user’s fine-grained interests.Then,the candidate news is fused into the user’s interest expression process,distinguishing the level of user preference activity,activating the user’s interests related to the current time,and ultimately improving the user’s representation effectiveness.Multiple comparative experiments were conducted on a open news dataset,and the results showed that the various performance indicators of this method were better than the baseline method,and experimental analysis further validated the rationality of the design.(2)A personalized news recommendation method based on semantic interaction and multi-interest collaboration.The method’s core includes a Semantic Interaction Enhancement(SIE)module and a Multi-Interest Collaboration(MIC)module.To address the issue of insufficient semantic correlation caused by independently encoding news titles and content,the SIE news encoding module strengthens the combination of news titles and content through parallel semantic interaction networks at the word and domain levels.Furthermore,a cross-selection network is designed to introduce the title into the content feature aggregation process,further enhancing semantic correlations and improving the representation effectiveness of news content.To address the issue of being unable to explicitly model a user’s multi-interest representation,the MIC user encoding module divides the user’s history into multiple interest clusters based on topic label,and represents the user’s preferences in different topics using cluster vectors.Through the interaction and collaboration between multiple interest vectors,the interaction between interests is simulated to enhance expression ability.Finally,combined with the user’s own and realistic factors,multiple interest vectors are adaptively aggregated into user representations,thereby improving the user representation effect.A number of comparative experiments are conducted on two open news datasets,and the results show that the proposed method outperforms the baseline method in various performance indicators.Additionally,the effectiveness of each designed component is further demonstrated through ablation experiments. |
