Bootstrapping And Weighted Information Gain In Support Vector Machine For Customer Loyalty Prediction

Prediction customer loyalty is an important business strategy for the modern telecommunications industry in the global competition. Support Vector Machine (SVM) is a classification algorithm that widely used to predict the customer loyalty. SVM in predicting customer loyalty has a weakness that affects the accuracy in the prediction. The problem is the difficulty of kernel function selection and determination of the parameter value. Large datasets may contain the imbalance class. In this study, bootstrapping method is used to overcome the imbalance class. In addition, datasets also contain some features that are not relevant to the prediction. In this study, we propose to use Forward Selection (FS) and Weighted Information Gain (WIG). FS eliminates the most irrelevant features and the computation time is relatively short compared to backward elimination and stepwise selection. WIG is used to weight the each attribute. In order to handle the selection of SVM parameters, we use a grid search method. Grid search method find the best parameter value by providing parameter value range. The experimental results from some combination of parameters can be concluded that the prediction of customer loyalty by using samples bootstrapping, FS-WIG and grid search is more accurate than the individual SVM.