| With the development of the live broadcast,the number of live users continues to grow.For Internet platforms which have both video and live broadcast service,because of the limited daily recommended content,it is necessary to accurately locate the content needs of users to make appropriate recommendations.Therefore,whether users will watch live broadcast has become a new issue for Internet platforms.Based on the real business problems,this paper studies the behavior of users who have watched game live broadcast.This paper takes the game live broadcast of Internet platform A as an example,takes users who watch the game live broadcast on the t day as the research object,and builds a model to predict whether users will continue to watch the game live broadcast on the t +1 day to help the platform locate needs of users.The main work is as follows:（1）In terms of data features,in order to collect more behavior information,features such as the behavior of watching videos are innovatively added.We also use text information such as profiles which are crawled from the platform and use text mining to build relevant features.（2）In terms of model methods,the DNN and Deep FM are used for prediction,and the effect is compared with machine learning models such as Random Forest,Light GBM,and XGBoost which use m RMR for feature filtering.The results show that the effect of Deep FM is better than other models.（3）Use the SHAP framework to analyze the feature importance of Light GBM.It is found that features such as watching live broadcast,watching related videos,age,time information and followed user information have an important impact on whether users will continue to watch the game live broadcast.This paper has a good effect on the prediction of watching game live broadcast.It adds new features and proves that some features have a good effect.It is feasible for the Internet platform to predict relevent problems,and it helps analyze factors which influence whether users will continue to watch live broadcast. |
