Research On Customer Churn Prediction Based On Ensembled Deep Learning

With the development of social economy,Enterprises are facing increasing competitive pressures,the market tends to be saturated,and customers have liquidity,many enterprises are facing the problem of customer churn,customer churn prediction(CCP)is particularly important to the development of enterprises,potential churned customers can be predicted in advance through CCP,CPP also can help enterprises to find the reason of customer churn,so that the enterprises develop targeted and efficient customer retention measures,the purpose is to effectively reduce the cost of enterprises and improve market competitiveness.and there are many solutions for customer churn prediction,traditional machine learning methods have the disadvantage that feature engineering has great influence on model effect,deep learning makes the algorithm less dependent on domain expertise and artificial feature extraction.There are some high-dimensional features and time series features in customer churn data,how to construct an effective deep learning model has become an important research topic in the field of CCP.In this paper,two customer churn prediction algorithms are proposed,it brings great improvement on customer churn prediction throught ensembled deep learning,and one application of customer churn prediction is constructed in this paper,The main research works are as follows:(1)A customer churn prediction algorithm based on covolutional neural network weighted ensemble(Swa CNN)is proposed.Currently,common customer churn prediction algotithms are relatively simple,the ability of feature extraction is limited,which cannot dig data deeply.Based on the above problems,Swa CNN is proposed to enhance model representation ability and build an integration model to improve prediction effect.The algorithm extracts features by stacking one-dimensional convolution layers,when use covolutional neural networks to predict customer churn,and also use stochastic weight average to obtain weighted ensemble model during the training process.Compared with single covolutional neural network model,Proposed method brings improvement on customer churn prediction.(2)A customer churn prediction algorithm based on Self-Attention LSTM is proposed.Traditional machine learning methods does not work well with time series features,it is of great significance to construct an effective model for churn prediction of customer data with time series features,long short-term memory(LSTM)is often used in time series features.However,a single LSTM model has insufficient generalization ability and insufficient attention to important features when dealing with complex problems,and it didn’t pay enough attention to important features.Combining Self-Attention with Bidirectional long short-term memory(Bi LSTM)to form SA-Bi LSTM model,similarly,combining Self-Attention with convolutional long short-term memory(Conv LSTM)to form SA-Conv LSTM model.Finally,the SA-BILSTM and SA-CONVLSTM models will be integrated to output the final prediction results.Proposed method brings improvement on customer churn prediction.(3)The competition among enterprises is fierce,and the important factor to improve the competitiveness is to reduce customer churn.It is of great significance to develop a customer churn prediction system so that the managers can take measures to retain customers in advance.Based on the proposed Swa CNN algorithms,a customer churn prediction application based on B/S framework is constructed.The application also has the functions of customer composition analysis and customer relationship management,which is convenient for operators to manage customers and take appropriate retention measures before customer churn.