Performance Evaluation of Various Classification Techniques for Customer Churn Prediction in E-commerce

It is always a challenge to predict the customers on the verge of churn accurately in e-commerce due to the complexity of features and dynamicity of data and develop effective churn prediction models to predict potential churners accurately. This paper presents an in-depth comparison between four machine learning techniques namely neural network, support vector machine, Naive Bayes and random forest, and Adam deep learning technique, for predicting customer churn in e-commerce. The classification techniques are implemented on benchmarked Brazilian e-commerce dataset. For the feature selection, principal component analysis and neighborhood component analysis techniques have been used. A balanced dataset, consisting of 11224 samples, is taken for study. The performance of the developed models is evaluated using the performance metrics viz. accuracy, sensitivity, specificity, true positive value, and true negative value. It has been found that the random forest classifier for the features selected using the neighborhood component analysis technique gives the highest prediction accuracy of 99.35% in comparison to classifiers used in this work as well as classifiers used by pre-vious researchers. Additionally, the accuracy of the classifiers for features selected using the neighborhood component analysis technique is higher as compared to the principal component analysis technique. In future, authors are working further to improve the performance of the developed model by incorporating more features as well as evaluation parameters and proposing new models using convolutional neural networks. The authors also intend to use more than one dataset for the training of the models in the future.