Abstract
This research introduces a hybrid model designed based on an explainable and data-driven paradigm of predicting customer loss in an online shop. The model incorporates Convolutional Neural Network (CNN), Bidirectional Long Short-Term Memory (BiLSTM), along with Attention, and Multilayer Perceptron (MLP) branch to reflect the presence of both time-based and space-based correlation in customer behavioral data. The experiment relied on an open-access Kaggle e-commerce dataset consisting of 16 attributes and 10,000 customer records. The Synthetic Minority Over-sampling Technique (SMOTE) is used to preprocess the data and balance, encode, and normalize it. The hybrid model created surpassed the traditional baselines, such as the Logistic Regression, Random Forest, and XGBoost, with an accuracy of 0.87, precision of 0.88, recall of 0.88, and F1-score of 0.50. Also, SHapley Additive Explanations (SHAP) analysis made interpretability a possibility as the key elements that caused churn were shown to be inactivity and support ticket frequency. The results can be discussed as the recommending model offer valuable information about proactive customer retention measures in e-commerce and enhance explainability, prediction precision, and robustness.
Keywords
Customer churn, Deep learning, Hybrid model, Explainable AI, E-commerce, SHAP