Prediction Model Of Consumer Churn In Big Data Environment

In recent years,with the rapid pace of life and the rapid development of mobile terminals,mobile games develop rapidly.Mobile games fill the fragmented time of mobile users.The market is of mobile games.However,although most companies accumulate a large amount of data,but did not cause enough attention,just stay at the stage of looking at the data,failed to dig the data in depth.The game's life cycle and profitability are closely related to its data analysis capabilities.With the increasingly fierce competition in the game market,on how to obtain greater benefits to extend the life cycle of the game,some prescient game companies choose to use big data to mine more and more detailed user groups to personalize and refine services.Data analysis plays a vital role in the operation and maintenance of the game,accurate data analysis is conducive to the introduction of reasonable novice guidance in the operation of the game,timely channel promotion and rich consumer scene design,which will greatly affect the attention of the game players,thus extending the life cycle of the game,and better profit.The churn prediction is very important in the mobile phone game product operation process,the enterprise according to the churn of early warning analysis results to develop retention strategies,so as to retain the core player groups of the game and extend the life cycle of the game.Predicting the behavior of players is of great significance to the development and management of game companies.The main contents and achievements of this paper are as follows:1.aiming at the problem that the feature dimension of mobile game data is too high,this paper compares three methods:principal component analysis,chi-square test and fisher ratio.Using logistic regression,support vector machine,decision tree,random forest and KNN method as the benchmark model,the results show that the main component analysis method will lose too much information,while the chi-square test and fisher ratio selection of feature correlation is high,fisher ratio can choose the best subset.2.aiming at the imbalance between active samples and lost samples in mobile game datasets,the methods of centralized under-sampling and oversampling are compared in this paper.An under-sampling method based on K-means clustering is proposed.The results show that the new method can significantly improve the prediction performance for a few samples.The combination of the new method and SMOTE method can also combine the advantages of the two methods,so that both the F value and ROC-AUC can be significantly improved.3.aiming at the problem of value difference of players,this paper proposes an improved AdaBoost model based on customer value,which can effectively solve the problem of value difference.