Research On User Visualization Analysis And Prediction Based On Data Mining

Under the background of mobile internet,the popularity of smart terminal applications has led to the rapid development of social media.Short video,In the context of mobile Internet,the popularity of intelligent terminal applications has led to the rapid development of social media.A short video is popular with the public because of its large information carrying capacity,fragmented dissemination and humanized content output form.Mobile short video social media has risen rapidly and gradually expanded to various fields,and people's lifestyles have undergone major changes with the development of short video industry.With the advent of intelligent Internet,the short video industry will enter a faster development stage.In order to adapt to the change of social development,enterprises change from "product-centered" to "user-centered".By analyzing this massive user data by means of data visualization,we can dig out the potential value of relevant users,help enterprises understand the behavior habits of existing users,and provide users with services that can meet their needs.At the same time,the user prediction task based on data mining technology is helpful for enterprises to mine target users,thus strengthening the business value of users to enterprises.Therefore,data mining technology has effective application value and practical guiding significance for user research.Based on the above background,the visual analysis method and data mining technology are used to complete the tasks of user analysis and modeling prediction.Firstly,based on the social media platform with abundant user data-Aauto Quicker,Python and other tools are used to process and analyze user attribute and behavior data,and three user research methods are used to complete the visual analysis of user portrait,access behavior and active retention.The goal of this work is to explore the characteristics of users and formulate accurate prediction strategies to assist decision-making.Secondly,we extract many behavioral characteristics such as user registration,login,video viewing and publishing,interaction,etc.,and build four classic active prediction models to evaluate the active state of users in the next week,and improve the performance of the models by about 2%through weighted fusion technology.Finally,we propose a more practical iterative optimization method to simplify feature engineering and increase the effectiveness of the model.After demonstration,the performance results of the model are improved by 8%,and the user activity prediction task close to the actual application scenario is realized.The research results of this paper are summarized as follows:(1)In this study,firstly,according to the behavior differences between abnormal and normal short video users,three user research methods are mainly used to visually analyze user attributes and behaviors from four levels,and the practical significance of active and retained users is introduced through multiple indicators,which provides research directions for the next experiment.(2)go deep into the previous research content,firstly,study the characteristics of shallow machine learning algorithm in data mining field,find the method of dividing data set by sliding window method,and build a huge feature project according to the characteristics of user behavior.then,after modeling by classical machine learning algorithm,compare and select four kinds of classification predictors suitable for this scene:logical regression,decision tree,LightGBM and shallow neural network,and carry out weighted fusion experiments on the latter two,and achieve certain results,but not very significant(3)In view of the complicated and time-consuming characteristic engineering tasks in the traditional user behavior modeling process,it is difficult to achieve accuracy and universality in practical business applications.In addition,considering the limitations of the shallow neural network model in time series data,it is decided to use GRU with gate mechanism to improve the standard RNN neurons.On this basis,combined with business needs,an iterative optimization method combining cosine fire reduction and hot restart technology is proposed to minimize errors and make the performance of this prediction model closer to the real results.The demonstration has made a great breakthrough in this experimental data set.