| The development of the Internet is accompanied by the problem of information overload.In order to solve the problem of information overload,personalized recommendation systems were built in mang fields to facilitate users to obtain information efficiently and optimize the user experience of the platform.For example,news information platforms often build recommendation systems based on user attributes,news attributes and user historical behavior.In the process of reading news information,users are recommended to read content that users are really interested in,which greatly reduces the time for users to retrieve information and brings more loyal users to news information platforms with efficient and convenient news reading experience.Therefore,recommendation click-through rate is an important index to evaluate the quality of recommendation system.The problem of News Click prediction based on text variables is mainly considered in this professional master dissertation.Firstly,three variables of user's historical reading news label are defined about a news App.Through the correlation analysis of the three variables,we get the types of news that users are interested in and their news reading behavior habits.Secondly,the influence factors of whether exposure news is clicked and read are studied by normalized mutual information method.Based on the results of mutual information analysis and association analysis,two variables are proposed to describe the relationship between user's historical behavior and current recommendation,including the combination variables generated by user's historical reading news label variables and exposure news label variables,and the intersection generated by association rules.Mutual effect variable.Finally,factorization machine(FM)model is used to establish click prediction models of exposure news with different combinations of variables.Accuracy,Precision and Area under curve are selected to evaluate the model,and the main factors affecting users' click-to-read exposure news are analyzed.It is found that:(1)The types of news that users are interested in mainly include entertainment,society,emotion,military,sports,funny,international,historical,pet and so on.Among them,entertainment news is the most popular type of news.(2)Users tend to read different types of news,and very few of them read only one type of news.(3)Among many types of recommended news,the click-through rate of policy news is low,while that of digital news is high.(4)Among many factors,user's reading history and current exposure types contribute most to predicting whether users click to read exposure news or not.(5)Knowing the relationship between user's reading history and current exposure news types is helpful for predicting whether users will click to read exposure news,such as whether users have read current exposure news types in the past,the possibility of reading current exposure news types according to users' latest reading history,and so on. |
PDF Full Text Request