A User Portrait Method Based On Multi-Source Weighted Fusion

The topic of this thesis is derived from the national key research and development project "Public Cultural Resources Intelligent Co-construction and Sharing and Management Platform Construction and Demonstration Application(2019YFC1521405)".With the rapid development of the Internet,users have left a large amount of behavioral data on the Internet.In the era of big data,user behavioral data is a valuable asset.Based on this,building user portraits is of great significance for precision marketing,advertising,and personalized recommendation.User interest prediction is a popular research direction for user portraits,and user behavior data is of great value for mining user interests.There are generally two problems with predicting user interests:(1)Analyzing the user's single type of behavior content data will cause inaccurate prediction results;(2)Analyzing the content data of all historical behaviors of the user will cause the amount of analyzed data to be too large,The analysis efficiency is low.In response to the above problems,based on the blogs involved in 6 different user behaviors:posting,favorites,likes,dislikes,comments,and browsing,this article conducts the following research:(1)Constructed a user portrait model based on Text CNN.Taking the blogs published by users as an example,a user portrait model of a single behavior data is constructed by using Text CNN to classify the text content of the blogs published by users.The user interest profile model based on Text CNN is used to test the user interest expressed by the 6 types of user behavior data,and the 4 types of data that can stably express the user interest are selected,namely: published,favorite,liked,and browsed blogs.As the source data for the subsequent prediction of user interest,the initial weight of the model is determined according to the experimental results.Since user interest changes dynamically,the user interest portrait model based on Ebbinghaus forgetting curve model and Text CNN test the stability and accuracy of the interest prediction results of the four types of user data collected in different time slices.Determine the time slice for collecting user data by comparing the stability and accuracy.Since historical data will affect the current user interest,the user interest profile model based on Text CNN tests the accuracy of the interest prediction results using user data in different numbers of time slices,and compares the accuracy to determine the current time slice and the previous one.Three time slices of user data are used to predict user interests.(2)On the basis of the above research,a user behavior data collection and weighting method for interest dynamic prediction is proposed.This method collects various types of data that can express user interests in different time slices,and determines different according to the importance of the time period.The ratio of the number of data collected in the time slice is determined according to the importance of the data type,and the ratio and weight of the number of collected data are determined to make the prediction result more accurate.Use Text CNN to weighted fusion of collected different types of user blogs to construct a user portrait model of weighted fusion of multi-source behavior data.(3)The user portrait model of weighted fusion of multi-source behavior data is optimized.The model data weights are determined through experiments,and the optimal user behavior data collection and weighting method and portrait model are obtained.The experimental results show that using the user behavior data collection and weighting method proposed in this thesis,the accuracy of prediction is improved by 6.8% compared to using a single behavior data;at the same time,the real-time changes of dynamic reflection of interest prediction results have also been greatly improved.There are 15 figures,37 tables and 52 references.