Churn Prediction using Attention Based Autoencoder Network

Abstract

User Behaviour Analysis gives valuable insights for customer management. Especially in Telecom sector it can help find the customer churn rate. In Telecom, customer churn is to find whether the customer is going to leave the service of the current operator or not. Effective churn rate prediction is a critical task. Classification models are often used for churn rate prediction. But most of these models have several shortcomings. They require manual feature extraction, or the model cannot balance skewed datasets. To overcome these problems, in this paper, we are proposing a feature extraction model based on Autoencoder with attention mechanism. The Autoencoder network represents the data in latent space representation. Attention mechanism makes the network focus on the features that highly contributes to the target prediction. The network contains a separate classification module for churn prediction. This technique results in using fewer training set yet producing superior results. The proposed model is compared with other classification algorithms and the results show that our proposed model outperforms other models in terms of accuracy and area under the curve.