Data Based Prediction Of APP User Beahavior

In recent years,smartphones have occupied the vast majority of the mobile phone mar-ket with their powerful features.As the major component of smartphones,mobile application(APPs)have also shown a booming development.In order to stand out in the fierce market competition,APP developers should think about what users think in order to meet the demands of users as much as possible and provide users with the most convenient service.With the de-velopment of data storage technology,a huge amount of user behavior data generated by APP users can be stored and the behavioral habits of users can be obtained and the needs of users can be obtained by in-depth mining and analysis of these data.We can even predict certain behaviors of users in advance,thereby providing decision support for the development and op-eration of the APP in order to provide users with more personalized services and enhance the user experience of the APP.For APP developers,user churn and payment are two behavior characteristics that they are most concerned about.Therefore,we carry out systematic research work baseed on APP users behavior data to study the prediction algorithm of user churn and payment.The main contributions of this paper are as follows.1.Data preprocessing.We first deal with the problem of data anomalies and duplication in the data set.Then,key statistics that reflect the user's behavior habits are extracted from the data set,and a simple user portrait is constructed.Also,We conducted some general analysis of the user's behavioral characteristics,and clarified the focus of subsequent user behavior.2.User churn prediction based on survival analysis.In order to solve the problem of data censoring,we use survival analysis to predict the user's survival function,which indicate probability of the user churn in the whole life cycle.By taking the median survival time as an estimate of the user's life cycle time,we have reached the goal of predicting when the user will churn.3.User purchase prediction based on ensemble learning We use ensembling learning to pre-dict whether users will pay or not in the future,and compare it with the prediction effect of some single models.The results show that the ensembled model has better results than the single model.