Design and Implementation of a System for Comparative Analysis of Learning Architectures for Churn Prediction

Abstract

Telecom companies are increasing their efforts in customer retention because acquiring new customers often costs much more than retaining existing ones. Therefore, it is important for operators to predict customer churns rapidly and accurately. Machine learning (ML) has been widely used for predictive churn modeling. However, classical ML methods require manual feature selection and time-consuming data preprocessing steps. To overcome these limitations, there is a paradigm shift toward deep learning (DL) for predicting churners. Although DL appears to be promising, the existing literature lacks comparative analysis of ML and DL techniques using benchmark churn datasets. Additionally, various DL architectures must be empirically evaluated to determine which type works best on churn data. We present a system for comparative analysis of learning architectures. Two benchmarked datasets, Cell2Cell and KDD Cup, serve as a use case of our system to provide insights on the extent of improvement DL can bring over classical ML. Four popular evaluation measures are used to compare the performance of popular DL architectures. Our experiments found that convolutional neural networks gave the best results in both use cases.