Explainable Customer Churn Prediction Model Based on Deep Learning

Customer churn represents a significant peril to the sustenance and expansion of business enterprises, necessitating proactive anticipation and effective management strategies by organizations. In response to the existing issue of poor interpretability in commonly used customer churn prediction models, this paper proposes an interpretable customer churn prediction model based on deep learning. The model first preprocesses the raw customer dataset and then applies three different sampling techniques—SMOTE oversampling, SMOTETomek hybrid sampling, and Tomek Links undersampling—to generate three new datasets. These datasets are subsequently fed into the TabNet model for training. The TabNet model executes multiple decision steps, where attention transfer modules are employed for feature selection in each step. Following feature transformation modules, the model learns advanced representations of customer features, using them as inputs for the attention transfer modules in the next decision step. The outputs of all decision steps are aggregated to obtain the final prediction result. Experimental validation on two publicly available datasets demonstrates that the proposed model outperforms other algorithms in terms of predictive accuracy and F1 score, while maintaining good interpretability.