Abstract: Taking advantage of machine learning algorithm in classification prediction, this paper explores a user churn prediction model of paid knowledge live through empirical research, analyzes the prediction variables, and provides decision-making basis for user retention management. Taking Zhihu live as data source, starting from two dimensions of user value characteristics and user review characteristics, users’latest consumption time, monthly average consumption times, average consumption amount, first consumption time, rating and comment text are collected, and then prediction models are constructed based on six different machine learning algorithms, and their prediction effects are compared. Then, the contribution of variables in the prediction of user churn is compared and analyzed. According to the key variables, the types of churn users are divided, and the corresponding retention strategies are proposed. Rating and comment sentiment have significant effect on user churn prediction; XGBoost model based on ensemble learning has the best performance, followed by random forest, so the superior generalization performance of ensemble learning has been well verified. By analyzing the important factors that affect user churn prediction, four types of churn users are summarized.

Key words: Machine learning, Knowledge online live, Paid knowledge, User churn, Prediction effect;User value, User evaluation