Sequential churn prediction and analysis of cellular network users-a multi-class, multi-label perspective

Abstract

We investigate the problem of churn detection and prediction using sequential cellular network data. We introduce a cleaning and preprocessing of the dataset that makes it suitable for the analysis. We draw a comparison of the churn prediction results from the-state-of-the-art algorithms such as the Gradient Boosting Trees, Random Forests, basic Long Short-Term Memory (LSTM) and Support Vector Machines (SVM). We achieve significant performance boost by incorporating the sequential nature of the data, imputing missing information and analyzing the effects of various features. This in turns makes the classifier rigorous enough to give highly accurate results. We emphasize on the sequential nature of the problem and seek algorithms that can track the variations in the data. We test and compare the performance of proposed algorithms using performance measures on real life cellular network data for churn detection.