An Ensembled DNN and XGBoost Model with SVM Augmentation in Churn Prediction

Customer churn prediction benefits the telecom company in retaining churning customers. Various methods, from traditional data analysis to advanced deep learning, were introduced in previous research, while there is little research on boundary data processing and combining advanced deep neural network architectures and traditional data-driven machine learning algorithms. This essay proposed the SDB model, selecting boundary data using SVM and making predictions using a stacked ensemble model where DNN and XGBoost made independent predictions, and logistic regression was used to fit the two predictions into a final prediction. Based on the dataset provided by Kaggle, the SDB model with various augmentation rates and several single models was trained to compare the accuracy, F1-score, recall and precision. Comparison showed that it obtained higher accuracy and F1-score than single models and an ensembled model without boundary augmentation. Meanwhile, the augmentation rate should be carefully chosen to control the trade-off between absorbing attention in boundary data and shifting data distribution, and more real-world datasets could be used on this model to assess its practical utility in the future.