APP Recommendation And Unloading Behavior Prediction Algorithm Based On Time Series Mining

In the last few years,with the development of mobile Internet,mobile applications(i.e.app)have played an unprecedented role in our lives.It has gradually become an essential part of people's social life,such as food,clothing,shelter,entertainment and work.For users,faced with the various applications in the market,they often get caught up in the dilemma of how to choose.For developers,predicting whether users will uninstall Apps in the future can help them find the defects of their products and improve their products.In addition,by anticipating whether a user will uninstall an App,the developers can recommend more high-quality similar apps in order to improve the user's recommendation experience in advance.The appearance of recommendation system has solved the problem of information overload.Traditional recommendation systems are generally based on the score,which recommend popular applications with high scores.However,the user's needs are often diverse and the objectivity of the application scores may not be guaranteed because of human manipulation.Therefore,the performance of this single recommendation method is often not satisfactory,which is also the significance of personalized application recommendation system.Besides,there are few researches on application uninstall prediction.In this paper,a personalized recommendation algorithm and app uninstalls prediction algorithm based on time series data mining are proposed.Time series refers to the observed dynamic data sequences arranged in chronological order.The data used in this paper is from AppChina-a leading APP market in China,which contains a dynamic usage log data of 200,000 users and more than 100,000 Apps-with typical time series characteristics.By analyzing App time series data of the users and extracting features,we can obtain the correlation between users and Apps.By using the information,a scoring model is established to get the users' ratings of the application,and then the recommended list is obtained by using the object-based collaborative filtering algorithm.At the same time,the appropriate classification model are selected according to these characteristics data to realize the prediction of the unloading behavior of the user Apps.In the recommendation experiment,we compares the traditional scoring recommendation system,and the results show that the personalized recommendation algorithm based on time series mining has higher accuracy.In app uninstalls prediction experiments,a variety of classic machine learning algorithms are compared including: SVM,Decision Tree and Naive Bayes,k-nearest neighbor algorithm,Logistic Regression,Random Forests,GBDT,etc.By comparing the prediction results of two sets of different characteristic data mining through time series,we found that the prediction accuracy was significantly improved after the user uninstallation ratio abstracted through time series were considered.In addition,GBDT,logistic regression and support vector machines usually have better prediction performance among these prediction models.