| Summary: | 碩士 === 銘傳大學 === 資訊管理學系碩士在職專班 === 92 === This research uses four constructs with 14 variables to build a customer churn prediction model. The constructs are customer classification, customer satisfaction, customer loyalty, and service quality.
At first, samples are collected through eliminating incomplete data by using the data mining technique. Then, the data is sequenced and statistically analyzed for its distribution. Furthermore, a prediction is given based on the resulting determining factor. Finally, the customer churn prediction model is built based on the above determining factor associated with data mining techniques. The prediction provides a goal, which helps the company to adjust their customer keeping policy, which will decrease customer churn rate therefore benefit the company.
This research is based on the customer data of an internet service provider (ISP) of Taiwan. The total effective data are 12177 records. A total of 10655 records, or 87.5 % effective data, are assigned as training data. Finally, the decision tree classification technique was applied onto the customer churn prediction model.
The result shows that customer loyalty is the determining construct of prediction customer churn rate. In addition, there are 9 major factors, which yield good prediction. The major factors are total used time, customer service record, discounts, application time, purchase amount, disconnection rate, date of most recent purchase, charge rate, and customer classification. Of the nine major factors, two of them are Categorical variables, and the remaining seven are numerical variables. In general, continuous numerical variables yields better prediction, because that ISP, like water supply company and electric company, provides services continuously. Customers behave differently as time changes. Therefore in order to predict a better customer churn rate, continuous numerical data should weigh more in the prediction model once a potential customer turned customer.
This research model is based on the analysis of traits of corporate ADSL subscribers. When the model applies on different products or industries, several types of data need to be fine tuned before they are collected, based on characteristics of the product or industry. After the data is collected, it will be filtered based on its predictability. Some variables does not yield precise prediction by themselves. In order to decide whether these variables are deciding factors, they have to be considered collaterately with their distributions.
The data, grouped by prediction that it is going to yield, is divided evenly among each group. The more diverse the data lies, the better prediction it yields.
|
|---|
