Development of fading channel patch based convolutional neural network models for customer churn prediction

Currently, Customer churn is a major challenge for e-commerce companies. It is necessary to have customer churn prediction model for e-commerce companies to predict the customer churn in e-commerce applications accurately. In this paper, a novel concept of fading channel patch-based heat map for the training of convolutional neural network deep learning models has been proposed. The objective of the present work is to train the basic, two-layered and three-layered convolutional neural network churn prediction models using benchmarked Brazilian e-commerce data heat maps. A pre-processed and balanced dataset containing 14,188 data samples of e-commerce customers is used for prediction. The models are trained using 70% of the data (9932 samples) and tested using 30% of the data (4632 samples). The heat maps, containing attributes and relevant purchase information for each customer, are generated and used for training of the developed models. The performance parameters viz. accuracy, lift, true positive rate, and false-positive rate are taken for model evaluation. The accuracy of the developed models in this work is also compared with the existing models developed by various researchers in the past. It is found that the two-layered convolutional neural network model has achieved higher accuracy and performance as compared to the three-layered and basic convolutional neural network models. The accuracy of two-layered convolutional neural network model is better as compared to existing machine learning and convolutional neural network models. Hence, this work proposes an accurate two-layered convolutional neural network churn prediction model in e-commerce. In the future, the authors intend to improve accuracy by using an ensemble convolutional neural network. Authors are also working further to train the developed models with more than one dataset.