Predicting Customer Satisfaction Under Severe Class Imbalance in Online Retail Data

Abstract

This study examines the prediction of post-purchase customer satisfaction in an e-commerce setting using order-level transactional and behavioral data. Customer ratings from one to five stars are grouped into three classes representing low, medium, and high satisfaction. The resulting classification task shows a strong skew toward high ratings. Three supervised learning models are evaluated: a support vector machine with a radial basis function kernel, multinomial logistic regression with spline-expanded numeric predictors, and a random forest classifier. Input variables include demographic attributes, order characteristics, browsing behavior, and delivery information. Model performance is assessed using accuracy, macro-averaged precision, recall, F1-score, ROC-AUC, and confusion matrices. The random forest achieves the highest accuracy by predicting the dominant class, while the support vector machine and spline-based logistic regression show lower accuracy with more balanced class-wise results. The findings indicate that distinguishing dissatisfied and moderately satisfied customers remains difficult with the available feature set and label design