K- local maximum margin feature extraction algorithm for churn prediction in telecom

Abstract

Telecom customer churn data is not publicly available because involving users' personal privacy. In 2009, the French telecommunications company Orange for knowledge discovery and data mining (KDD) competition provides a telecom customer churn data set KDD Cup 09. In order to solve the high dimensional problem of KDD Cup 09, a new feature reduction method is used to explore the influence of different features on the prediction of classification model. In this paper, a new K- local maximum margin feature extraction algorithm (KLMM) is proposed. Through researching on the diversification subspace partition rules, the corresponding potential field structure is constructed. According to the data source in the dimension of scalability, the intrinsic link between data attributes and classification results is revealed. The extracted features can reduce the dimension of the churn prediction in telecom data. The KLMM method adapts auto selection sigma factor to reflect the anisotropy of features. The potential function is used to assess the weights of attributes and find the potential important weight. Experiments and analysis show that the extracted features by KLMM are more likely to find a classification hyperplane which can separate data points of the different classes.