Feature Subset Selection Using IULDA Model for Prediction

摘要

With the decrease in tariff rates and growing popularity of telecom options, competition in the field for abstracting customers and expanding market is becoming fiercer. It is evident by research that the cost levied by losing a customer from the telecom affiliation is sixfold more drastic than the profit that of adding a new one. The proposed Indexed Uncorrected Linear Discriminant Analysis (IULDA) classification model for customer churn prediction effectively handles increased amount and dimensionality of data and has been tested on L-class problems of UC Irvine Machine Learning Repository and real dataset of the train sample-5,200 customers, the calibration sample-3,680, and the test sample-4,500 observations. The objective evaluation of the investigated methods was measured by precision, specificity, sensitivity, and accuracy by implementing the MATLAB tool. The accuracy of the IULDA model was 95% for UCI churn datasets and 72.4% for real customer datasets, respectively.