Research On User Profile Construction And Community Recommendation Technology

With the blossom of internet applications,news recommendation system is the most efficient tool for users to acquire information from thousands on thousands news.Let users quickly obtain effective and diverse news from the mass of news and satisfy the need of the users’preference is the main task of news recommendation system.As an important application of natural language processing（NLP）,news recommendation system has become a research hotspot in industry and academia.It is significant and important for industry solving actual problems to research on news recommendation system.News recommendation system could reduce the cost of users and promote the efficiency of reading.The research on news recommendation system will be referenced by other fields of recommendation task.Consequently,research on news recommendation has become a very fatal research work in the field of recommendation.By analyzing the characteristics of news,this thesis proposes a user profile and community recommendation scheme based on the form of tag to solve the existing problems in the field of personalized news recommendation,in which include the poor interpretability of the recommendation model,inconvenience caused by massive data,the low diversity of recommendation results,and the inaccurate expression of user interests.（1）In order to solve the problem of poor interpretability of recommendations,the thesis designs an improved algorithm for extracting news tags based on the existing news representation methods and the characteristics of news writing,and it proposes optimized strategies for different channels of news.Tag-based news recommendation can produce intuitive explanations for recommendation results.（2）In order to ease the problem of cold-start in news recommendation system,the thesis proposes a research on community recommendation methods.Based on the idea of"people are divided into groups",the thesis uses the definition of knowledge graph to represent users and generates groups through clustering algorithm,then calculates effective recommendation results for users based on community profile.Community recommendation weakens the phenomenon of cold-start problem in the recommendation system.（3）In order to solve the problem of low diversity of recommended results,the thesis proposes a secondary user profile structure in the user profile construction stage.The secondary user profile is clustered for producing users’preference topics,and according to the proportion of the topic of tags,the user recommendation results are calculated.The experimental results prove that the hierarchical structure of second-level user profile has an improved effect on the diversity of news recommendation results compared with the first-level user portrait structure.（4）In order to solve the accurate representation of user preference in news recommendation system,the thesis proposes the strategy for updating and attenuation of user profile Based on the form of tag representation,the cycle of attenuation for different types of news is counted,and the attenuation formula of user preference is designed to make the profile dynamically shift with user interests.The experimental results show that the updating and attenuation of user profile could track the dynamic changes of user interests and improve the accuracy of the recommendation results.The method proposed in this thesis is experimentally verified on the online news data.Compared with the tool of news tagging from original company platform under the same conditions of manual evaluation,it finds that the accuracy of this tagging method is 34.9%of the original ones.The user profile module has carried out experiments and verifications on the construction,attenuation and update operation and the selection of parameter.The result of community recommendation is evaluated by two ways,which is the calculation of P,R,and ₁F,involving the assessment by artificial ways.Based on the comparison of the two kinds of assessment method,the experimental results prove that the proposed user profile structure and community recommendation method really have high accuracy and diversity in news recommendation task.