Churn Prediction Model Based On Wide & Deep Learning

With the popularity of the 4G network and the rapid development of the Internet,telecom operators have launched a variety of promotions to attract users together with other companies.These activities have attracted more and more new users in the network and also causes the accelerated depature of old users from the network.This phenomenon of accelerating off-line has attracted widespread attention in the industry.How to accurately make churn prediction before users leave the network,so as to take a series of measures to retain the users and reduce the loss of the operator has become a focus of research.To solve the churn prediction problem of telecommunication users,the industry has experimented with a series of machine learning algorithms.Due to the rapid development of the deep learning technologies,more and more researches have utilized neural network models in the task of churn prediction.To solve the problems existing in churn prediction models based on neural network,we design and construct a neural network framework based on wide & deep learning in this paper which is based on the real user data provided by telecom operators.The work of this paper mainly has the following three contents:1)This paper proposes a churn prediction framework based on wide & deep learning.Linear model with cross product transform can effectively memorize the co-occurrence relationships between features,while neural network can automatically generate a large number of complex feature combinations to get better generalization capability.However,when the neural network's input contains discrete features,there may be an issue of overgeneralization.So we learn the idea of wide & deep learning in churn prediction.The linear model is used as wide model and the neural network is used as deep model.The two parts are combined and trained together to combine the benefits of linear model and neural network.Experiments results show that the performance of the neural network with linear model can be increased by up to 5.39% on PR-AUC compared with the neural network without linear model.2)At present,the layers of the neural network models applied to the churn prediction are relatively shallow,and can fit data well only when the amount of the data is small.However,the use of shallow models will lead to serious under-fitting problems as the amount of data increase.To solve the problem,we propose a deep convolutional neural network structure and attempt to add a short link or gate mechanism in the convolutional layers to fit a large amount of data.Compared with the shallow network as baseline,the performance of our deep network can be increased by up to 72.33% on PR-AUC.3)The existing data-driven churn prediction models usually select only one kind of time granularity to aggregate time series features and build models on static features and time series features using machine learning algorithms.Such approaches only consider the impact of the models on classification performance,and do not fully consider the role of data.In this paper,we attempt to use monthly and daily time granularity to aggregate features by feature-levle fusion and decision-level fusion,and make a series of model training and model ensemble based on the extracted features to further improve the performance of the model.Experimental results show that the performance of our method which simultaneously uses feature-levle fusion and decision-level fusion can be increased by up to 21.94% on PR-AUC.