Research on User Churn Warning based on Machine Learning

In the current competitive communication industry, how to avoid user churn has become an important issue for enterprises. The development of big data technology has provided a new way for communication companies to predict subscriber churn. This paper predicts user churn based on 900,000 data from QD Mobile using a dataset processed by random sampling. The accuracy of three algorithms, including decision tree, random forest and AdaBoost classifier, is compared for user churn prediction. The random forest algorithm is found to be the most accurate for user churn prediction. Based on this, grid search algorithm in machine learning is used to find the best parameters of the random forest model and improve the prediction accuracy to 81.84%. The result can help communication companies to predict subscriber churn probability and improve competition level.