Graph Learning Algorithm For Tele-com User Churn Prediction

With the popularization of smart phones in the era of mobile Internet,the development of mobile phone business including telecom is getting faster and faster,and there are more and more telecom competitors.It is necessary to design an excellent user churn prediction algorithm.While the traditional user churn prediction algorithm relies on the experience of traditional feature engi-neering and ignores the change of users' product.Algorithm for graph relies on the detailed call records,which is illegal to privacy protection.We propose a automatic feature intersection and an online forecast algorithm based on temporal graph.Our method is better than traditional machine learning method and classic features intersection methods on offline datase.Besides,on unicom online dataset our online method is better than other methods.Besides,a comprehensive off-line and online system for user churn prediction is built,and the effectiveness of collaborative inference is proved in experiments.The main innovation points of this paper are as follows,1.Propose the structure of automatic features intersection based on graph algorithm frame-work,consist of local features intersection module and multi-head attention feature inter-section module and global graph feature intersection module.They can gengerate the zero order and 2 order,high order crossover feature,2.Propose a online update algorithm,a temporal graph algorithm based on multi-head atten-tion and memory mechanism.After the node embedded representation of the graph repre-sentation layer is obtained,the node memory vector is updated through the message-memory mechanism,and the updated results can be directly used for the prediction of the next mo-ment.3.Propose a comprehensive collaborative solution for off-line and online user churn prediction,which effectively combines the advantages of offline model to construct massive features and online model to timely capture short-term trends.