A Study On Churn Prediction Of Fixed-Line Subscriber

With the growing substitution by mobile communication services and increasing competition in the fixed-line market, domestic fixed-line operators are facing great challenges. The increasing loss of subscribers is one of the biggest challenges. The huge loss caused by the switch of subscribers and the great difficulty of winning new ones make the fixed-line operators realize the importance of subscriber churn prediction and subscriber retention. In response to the fixed-line operators' strong desires for churn prediction and the lack of researches and practices in the fixed-line market, this thesis studies how to apply data mining theories and technologies to churn prediction of fixed-line subscriber.Applying the CRISP-DM methodology, and combining it with the understandings of fixed-line business, this thesis elaborates the steps of building churn prediction model for fixed-line subscriber, including business understanding, data understanding, data preparation, modeling and evaluation. This thesis also points out the key issues of churn prediction of fixed-line subscriber, after summarizing the problems of the available data for churn prediction of fixed-line subscriber.The construction and selection of characteristics has great impact on the learning efficiency, accuracy and stability of final models. After analyzing various variable correlation theories, the thesis introduces ROC curves and mutual information theories to work out the method for characteristic selection. In the method, ROC curves are firstly applied to detect and deselect ineffective characteristics. Subsequently, mutual information is used to detect strongly correlated characteristics, among which characteristics with superior predictive performance are kept. Modeling method is the key to the effectiveness of prediction results. This thesis proposes mSTree-Logistic model, being inspired by TreeLogit model. In this model, a logistic regression function is induced from multiple decision trees, which are built based on different training sample sets respectively.A practice of churn prediction is conducted in a filiale of a fixed-line operator. The theories and methods proposed in this thesis are proved to be feasible and effective.