A New Churn Prediction Model Based on Deep Insight Features Transformation for Convolution Neural Network Architecture and Stacknet

Jalal Rabbah, Mohammed Ridouani, Larbi Hassouni

OAI: oai:igi-global.com:300342 • DOI: 10.4018/ijwltt.300342

Predicting churn has become a critical issue for service providers around the world. In particular, telecom operators for whom acquiring new customers is four times more costly than retaining existing ones. To keep up with the market, considerable investments are made to develop new anti-churn strategies, including machine learning models that are increasingly used in this field. In the proposed work, we combine three stages. In first stage, by using deepInsight, we transform the attributes of dataset into images in order to take the advantage of the strength of convolution networks in detecting hidden patterns in the dataset. In second stage, we use deep convolutional neural network for features extraction. In the last stage, we built a three-layer Stacknet of eight selected algorithms using a successive split-grid search for classification and churn prediction. The proposed model obtained the best accuracy score of 83,4%, better than the others proposed models in literatures.

Web-Based Teaching & Learning Education Educational Technologies