User Churn Prediction Of Video App Based On Behavior Data Mining

After nearly two decades of development,China’s online video industry has entered the stock competition phase.The rising cost of customer acquisition makes video platforms gradually take improving the retention value of existing customers as the core of their user operation strategy,and the user churn prediction is of great importance in improving user retention.The investigation of mining useful information from the massive behavior data of users and constructing a user churn prediction model is of great significance for video platforms.Based on the users’ behavior data of a video APP,this paper uses a variety of machine learning models and deep learning models to predict user churn.Firstly,based on the users’ log data,features are extracted from three dimensions:login behavior,video watching behavior and user interests,and four machine learning models(Random Forest,XGBoost,LightGBM,CatBoost)are used for training and parameter tuning.The results show that XGBoost is better than other models in AUC,recall and flscore.Secondly,the time-series features related to users’ behavior are extracted from the log data,and deep learning models based on GRU and CNN are used to make prediction based on uses’ time-series features and user profile features.The results show that both the CNN-based model and the fusion model of GRU and CNN(DLCNN and GRU-CNN)outperform the machine learning models.The GRU-CNN model effectively combines the advantages of GRU and CNN by parallel fusion,and can learn features automatically from uses’ time-series features more adequately,which performs best among the deep learning models.After adding the attention mechanism into the GRU-CNN model,the performance is further improved,and the AUC,recall and fl-score is increased by 0.7%,13.7%,5.8%respectively compared to XGBoost.Finally,the variables with the highest importance can be obtained according to the importance ranking of the features output by the model.Combined with the SHAP framework,it is found that features such as user interests,login behaviors,video watching behaviors and device memory have important influence on whether users churn or not.This conclusion can also provide some reference for the video platforms to formulate their user operation strategies.