Adaptable Churn Prediction Pipeline for Hybrid Business Model Using Deep Neural Networks and Gradient Boosting

Abstract

In the dynamic landscape of hybrid businesses, where contractual B2B-like and B2C-like customer relationships converge, the demand for effective churn prediction mechanisms is crucial. This study presents a predictive pipeline on real-life data, employing Deep Neural Network (NN) and Gradient Boosting (XGBoost)-based models for an Industrial Computer Business with such characteristics. Our findings demonstrate that the NN approach showed the ability to make more logical decisions. This claim is supported by utilizing the Euclidean distance metric, specifically applied to measure the similarity between incorrectly predicted instances and those belonging to the misclassified class. Moreover, to address a possible class imbalance scenario, we implement under-sampling with Tomek Links, improving the model’s robustness further. The developed models show impressive performance as the mean F1 score, calculated across multiple years, reaches an impressive 0.85, surpassing the 0.80 accuracy threshold. Furthermore, our deployment strategy offers a real-world example of integrating Robotic Process Automation (RPA) and AI. This approach not only showcases the synergy between RPA and AI in building a digital solution but also serves as a practical blueprint for incorporating the adaptability of predictive models in dynamic business environments.