| With the rise of the Internet and changes in the way people read,digital reading has developed rapidly.Online news platforms such as Sohu News and Netease News have attracted a large number of users to read digital news,but it is difficult for users to choose the news they are interested in from a large number of news every day.Personalized news recommendation technology can help users to filter and filter information and provide more valuable news content.The key to personalized news recommendation technology is to accurately understand and capture the semantic information of news texts to measure the similarity and correlation between different news.Currently,representation-based matching personalized news recommendation methods usually represent news text and user interests as vectors,and make recommendations by calculating the similarity or matching degree between them.However,this approach fails to fully consider the interactions between word pairs,thus limiting the expressiveness of the model.Although this method can capture the overall semantic information of news,it ignores the fine-grained relationship between words.Therefore,in some cases,the representation-based matching personalized news recommendation method may not be able to make full use of the interaction information of word pairs,resulting in a lack of accuracy and personalization in the recommendation results.On the other hand,analyzing the user’s historical interaction sequence and extracting the user’s interest features is also a very important step in personalized news recommendation.However,most current personalized news recommendation methods often do not consider the interaction with candidate news when modeling user interest feature representations.The interaction between candidate news and user interests can provide more contextual information,which helps the recommender system to better understand user interest preferences.If this interaction is ignored,the expressive ability of the model will be limited,and the user’s interest features may not be fully mined,resulting in degraded performance.Therefore,fully mining the semantic features of the news text and extracting the user’s interest features from the user’s historical interaction sequence is a meaningful research work for more personalized news recommendation.Through these works,the accuracy and user experience of the recommendation system can be improved,meet the personalized needs of users,and provide them with more valuable news content.This thesis comprehensively and specifically analyzes the personalized news recommendation task.At the same time,in view of the shortcomings of the existing news recommendation methods based on representation matching,this thesis proposes the following two solutions:(1)We propose a user interest-activated recommendation method that fuses multi-channel information.Contains two frameworks(interactive framework and distributed framework).In the interactive framework,a user multi-channel interest modeling framework MIF is proposed to capture more semantic clues related to user interests;in the distributed framework,a candidate-aware interest activation module TAR is designed,using different candidate news Vector to adjust the user representation learned from the user’s historical reading records,so as to accurately match the candidate news with the user’s interest part related to the candidate news;finally,effectively assign the weight of the two module scores,so that the model can better Fusion is carried out,and the effectiveness of the model in news recommendation is verified on the real data set of MIND news.(2)We propose a framework MnRec for fusing multi-granularity information for news recommendation.The two matching methods are fused by interactive attention and representational attention,and Bi-LSTM is used to learn the representation of each news,and an attention mechanism is introduced in the aggregation process to characterize the importance of each word to the news representation.Since users’ interests are multi-level and diverse,we finally designed the RTCN module to model users’ multi-level interests.Experimentally validated on the MIND dataset,the model achieves substantial improvements in prediction. |
PDF Full Text Request