Research On User Churn Prediction Of Communication Network Based On Spark Platform

In recent years,with the rapid development of mobile communication technology,the number of mobile communication network users increased dramatically,the com-munication market is close to saturation,and the competition among telecom operators is becoming more and more intense.At the same time,nearly saturated market makes operators pay more attention to the problem of user resource loss.For the communica-tion operator,it is possible to predict the potential loss of users by using the various data generated when the mobile user uses the mobile terminal,and to retain these potential users so that the market share and profit can be maintained.Therefore,the study of user churn prediction is of great significance to the communication operators.In this thesis,the customer churn prediction problem is studied from the two as-pects.One is the training speed of neural network algorithm and the other is features selection.First,we study the BP neural network algorithm.Back Propagation(BP)neural network algorithm has two weight update strategies,individual update and Full-Batch update.For the Full-Batch BP neural network algorithm,the algorithm is time consum-ing,We need to calculate all the samples in the dataset to update the weight,but the algorithm can be implemented in parallel.For the Individual BP neural network algo-rithm,we only need to calculate one sample to update the weight,so the weight updates quickly,but it can not be implemented in parallel.Therefore,we proposed Mini-Batch BP neural network distributed algorithm.By combining Full-Batch and Individual two weight update strategies,Mini-Batch BP neural network algorithm can improve the per-formance of neural network algorithm.Experimental results show that compared with Full-Batch BP neural network algorithm,the training time of Mini-Batch BP neural net-work is greatly reduced without losing the accuracy of prediction.Next,we discuss the value of the parameter K of the Mini-Batch BP neural network algorithm,and find that the value of K has a great influence on the training time.Then,the feature selection of user churn prediction problem is studied.We extract seven features,six user call behavior features and one user correlation feature.User relevance feature is the impact of the loss of the customer to his neighbors.We use the Spreading Activation Algorithm(SPA algorithm)to extract this feature.Subsequently,we used the first six call behavior features and all the seven features as training set to train and predict.We find that,after adding a user correlation feature,the prediction performance will be improved.Next,we confirm the role of user relevance from t-wo aspects:correlation statistics and the relative importance of features.Finally,we proposed a fast and accurate model for churn prediction based on two research points.The Mini-Batch BP neural network distributed algorithm used in this thesis can accelerate training and predicting processes,help to quickly and timely predict the loss of users.At the same time,the features of user correlation can improve the prediction accuracy effectively.Therefore,the fast and accurate prediction model proposed in this thesis is of great significance to the practical application of the user's prediction.