Research On Telecommunication Customer Churn Prediction Based On Recurrent Neural Network

The telecommunications industry plays an important role in every country,and there is also fierce competition.Since entering the 21 st century,China's telecommunications industry has expanded rapidly,and has penetrated into people's lives in many business fields such as mobile phones,fixed phones,and network lines.With the saturation of the market in recent years,the problem of telecommunications customer churn has attracted more and more attention from telecommunications operators.In fact,the churn of telecommunications customers has always been a key area of focus for the telecommunications industry.Among them,the loss of mobile phone customers is the most serious and most concerned because of the mobile nature of them.In our country,the reason for the continuous concerns of telecommunications customers churn can be explained in two aspects: On the one hand,as China's mobile phone market has truly reached a saturation stage,national policies have guided the improvement of mobile number portability strategies,which has greatly improved the convenience of customers to change telecom service providers.It will definitely induce a higher churn rate.On the other hand,telecom operators' data collection dimensions have increased and the frequency of collection has become higher and higher.Traditional modeling and prediction methods may not be able to fully mine the data.It is necessary to propose a modeling method adapted to the current data volume and data dimensions.At present,the methods of customer churn prediction in academia are mainly concentrated in several directions: a specific sampling strategy for data imbalance combined with traditional classifiers;using classic survival analysis methods to model customer churn from the perspective of survival data and analyze the impact of different factors on customer survival time;giving different weights to a variety of typical classifiers to improve prediction ability;using deep learning models with multiple hidden layers to extract as many useful features as possible for modeling.However,these methods have some disadvantages:(1)they cannot cope with the granularity of the current data and the dynamic data with time series characteristics;(2)they require complex manual feature engineering and the process is not intelligent enough.This paper starts with the construction of a survival analysis model and proposes an end-to-end Time-Varying Deep Learning based Survival(TVDLS)model.First use standard recurrent neural network to encode dynamic data,instead of tedious feature engineering,which can automatically and more fully extract dynamic data information.Next,a gated recurrent unit network(GRU)is used to build a TimeVarying Deep Learning based Survival model,where the loss function has been modified in the survival analysis model,which can have a better learning effect than the classic loss function.The model in this paper is compared with the previous methods at two levels:(1)comparing the survival function fitting ability with the classic survival analysis model and the deep survival analysis model,and using the two indexes of C-index and ANLP to compare,both get better results;(2)Comparing churn prediction ability with feature engineering and machine learning methods,and use the AUC and MP indicators for comparison.The AUC index can get better results.Finally,this paper conducts a subset analysis,considering several different customer data timing changes,and analyzes whether the model can fully learn such changes.The experiments show that the model in this paper performs better,which can help business personnel make more accurate predictions on specific customer groups.